A tower foundation deformation monitoring method and system coupling SAR and optical fiber

Abstract

The application discloses a tower foundation deformation monitoring method and system coupling SAR and optical fibers, and relates to the technical field of intelligent monitoring of power facilities, comprising: acquiring SAR images and optical fiber sensing data of a monitoring area, respectively performing time series interference processing and space-time alignment standardization processing, and constructing a multi-source standardized data set; inputting the multi-source standardized data set into a pre-trained double-flow neural network model, extracting macroscopic deformation space features and microscopic mechanical response features, and performing fusion; comparing the fused macroscopic deformation space features and microscopic mechanical response features with a pre-set failure mode feature knowledge base, outputting a failure mode and a confidence probability; dynamically adjusting a tower risk level according to the failure mode and the confidence probability, and generating a monitoring strategy instruction. Through deep feature fusion and dynamic closed-loop feedback, early-stage, high-confidence intelligent monitoring and early warning of tower foundation deformation, especially concealed erosion risk, are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for power facilities, and in particular to a method and system for monitoring the deformation of tower foundations coupled with SAR and optical fiber. Background Technology

[0002] In the operation and maintenance of ultra-high voltage transmission lines, accurate deformation monitoring of tower foundations in complex environments such as deserts and wastelands is crucial to ensuring power grid safety. The main approach is to combine space-ground integrated monitoring with spaceborne synthetic aperture radar time-series interferometry and fiber optic sensing technology. The former can acquire millimeter-level surface deformation over a wide area, providing macroscopic stability assessment; the latter can measure the microscopic physical parameters of the tower foundation itself, such as strain, temperature, and vibration, with high precision. Combining the two to achieve macroscopic and microscopic complementarity has become an industry consensus. The integration of existing technologies usually remains at the post-data fusion level, that is, the two types of data are processed independently and then spatially superimposed and manually compared. Post-event correlation analysis is performed on significant deformations that have already occurred, making it difficult to identify the early and proactive development process of potential hazards.

[0003] The core challenge facing existing technologies is to achieve the leap from data presentation to mechanism interpretation. This is particularly evident in the insufficient early identification capability of specific hidden risks such as tower foundation erosion. The initial signals of potential hazards are weak and have both acoustic and geometric characteristics. Conventional static, single-source threshold alarm mechanisms are unable to reliably extract effective signs from environmental noise. Existing systems lack a risk perception-based, dynamically adaptive intelligent closed loop, and cannot dynamically adjust monitoring strategies based on initial risk signs to achieve focused observation and targeted verification of potentially high-risk targets. This results in passive and delayed system response, making it difficult to complete the intelligent evolution from phenomenon monitoring to proactive early warning. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a tower foundation deformation monitoring method coupled with SAR and optical fiber to solve the core problems of existing technologies, such as difficulty in achieving the transition from data presentation to mechanism interpretation, insufficient early identification capability of hidden erosion risks, and lack of risk-based dynamic adaptive closed loop, resulting in passive and delayed response.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for monitoring the deformation of a tower foundation coupled with SAR and optical fiber, which includes acquiring SAR images and optical fiber sensing data of the monitoring area, performing temporal interferometry processing and spatiotemporal alignment standardization processing respectively, and constructing a multi-source standardized dataset. The multi-source standardized dataset is input into a pre-trained dual-stream neural network model to extract macroscopic deformation space features and microscopic mechanical response features, and then fuse them. The fused macroscopic deformation space features and microscopic mechanical response features are compared with a pre-set failure mode feature knowledge base to output the failure mode and confidence probability. The risk level of the tower is dynamically adjusted based on the failure mode and confidence probability, monitoring strategy instructions are generated, the sampling frequency of specific channels is increased, and computing resources are pre-allocated to special analysis algorithms. Real-time spectrum analysis is performed on the improved sampling frequency to identify specific characteristic spectra. The surface deformation sequence in the multi-source standardized dataset is re-analyzed to detect nonlinear trends and trigger special early warnings when matching spatiotemporal correlations. By integrating failure modes and confidence probabilities with specific early warnings, a comprehensive diagnostic report is generated.

[0007] As a preferred embodiment of the tower foundation deformation monitoring method coupled with SAR and optical fiber described in this invention, the method includes the following steps: acquiring SAR images and optical fiber sensing data of the monitoring area, performing temporal interferometry processing and spatiotemporal alignment standardization processing respectively, and constructing a multi-source standardized dataset: Acquire SAR images and fiber optic sensing data of the monitoring area, perform temporal interferometry on the acquired SAR images of the monitoring area, and obtain the surface deformation time series, deformation rate map and phase gradient field. Spatiotemporal alignment and standardization are performed on the acquired fiber optic sensing data of the monitoring area to obtain spatiotemporally aligned fiber optic sensing data. By integrating surface deformation time series, deformation rate map, phase gradient field and spatiotemporally aligned fiber optic sensing data, a multi-source standardized dataset is constructed.

[0008] As a preferred embodiment of the tower foundation deformation monitoring method coupled with SAR and optical fiber described in this invention, the method includes the following steps: inputting a multi-source standardized dataset into a pre-trained two-stream neural network model, extracting macroscopic deformation spatial features and microscopic mechanical response features, and fusing them. The macroscopic deformation feature extraction branch extracts the phase gradient field from the multi-source normalized dataset and inputs it into the pre-trained two-stream neural network model; the microscopic mechanical response feature extraction branch extracts the high-density strain field from the multi-source normalized dataset and inputs it into the pre-trained two-stream neural network model. The macroscopic deformation feature extraction branch based on the pre-trained two-stream neural network model processes the phase gradient field to obtain the macroscopic deformation spatial features; Micromechanical response feature extraction based on a pre-trained two-stream neural network model is used to process high-density strain fields and obtain micromechanical response features. The attention mechanism is used to fuse macroscopic deformation space features and microscopic mechanical response features to obtain fused multimodal features.

[0009] As a preferred embodiment of the tower foundation deformation monitoring method coupled with SAR and optical fiber described in this invention, the method includes the following steps: comparing the fused macroscopic deformation spatial features and microscopic mechanical response features with a pre-set failure mode feature knowledge base, and outputting the failure mode and confidence probability. Based on the contribution weights of each dimension in the fused multimodal features to the failure mode discrimination, a feature dimension weight vector is formed. The feature dimension weight vector is used to perform weighted similarity matching calculation on each failure mode feature in the fused multimodal features and the pre-set failure mode feature knowledge base to obtain a set of weighted similarity values. Select the failure mode feature corresponding to the maximum value from the weighted similarity value set as the matching failure mode feature, and map the matching failure mode feature to the failure mode. The maximum value in the weighted similarity set is converted into a confidence probability using a normalization function, and the failure mode and confidence probability are output.

[0010] As a preferred embodiment of the tower foundation deformation monitoring method coupled with SAR and optical fiber described in this invention, the method includes the following steps: dynamically adjusting the tower risk level based on the failure mode and confidence probability, generating monitoring strategy instructions, increasing the sampling frequency of specific channels, and pre-allocating computing resources for specialized analysis algorithms: The failure mode and confidence probability are input into the preset risk level assessment rules to obtain the updated tower risk level; Based on the updated tower risk level, a monitoring strategy instruction is generated; the monitoring strategy instruction is executed to increase the sampling frequency of the distributed acoustic wave sensor (DAS) channel related to the failure mode in the fiber optic sensor network monitoring data to a preset value; The monitoring strategy instructions are executed, and specialized analysis algorithms that identify erosion characteristic spectrums are pre-allocated computing resources.

[0011] As a preferred embodiment of the tower foundation deformation monitoring method coupled with SAR and optical fiber described in this invention, the method includes the following steps: real-time spectrum analysis of the improved sampling frequency to identify specific characteristic spectra, re-analysis of the surface deformation sequence in the multi-source standardized dataset to detect nonlinear trends, and triggering a special early warning during spatiotemporal correlation matching: Real-time spectrum analysis was performed on the distributed acoustic wave sensor (DAS) data after the sampling frequency was increased, and the real-time spectrum analysis results were obtained. The specific characteristic spectrum is identified by using real-time spectrum analysis results. The surface deformation sequence in the multi-source standardized dataset is reanalyzed to obtain the surface deformation sequence reanalysis results; Nonlinear trends are detected from the reanalysis results of surface deformation sequences, and the detected nonlinear trends are obtained. A special early warning is triggered when the identified specific feature spectrum matches the detected nonlinear trend in a spatiotemporal correlation.

[0012] As a preferred embodiment of the tower foundation deformation monitoring method coupled with SAR and optical fiber described in this invention, the method involves fusing failure modes and confidence probabilities with specific early warnings to generate a comprehensive diagnostic report, including the following steps: By correlating and aligning failure modes and confidence probabilities with specific early warnings in the spatiotemporal dimension, multi-source diagnostic information after correlation and alignment is obtained. The multi-source diagnostic information, after being integrated and aligned according to the preset report template, is used to generate a draft diagnostic report. The draft diagnostic report is formatted to output a comprehensive diagnostic report containing mechanistic diagnosis, risk level, and treatment recommendations.

[0013] Secondly, the present invention provides a tower foundation deformation monitoring system coupled with SAR and optical fiber, including a processing module, which acquires SAR images and optical fiber sensing data of the monitoring area, performs temporal interferometry processing and spatiotemporal alignment standardization processing respectively, and constructs a multi-source standardized dataset. The comparison module inputs a multi-source standardized dataset into a pre-trained dual-stream neural network model, extracts macroscopic deformation space features and microscopic mechanical response features, and fuses them. The fused macroscopic deformation space features and microscopic mechanical response features are then compared with a pre-set failure mode feature knowledge base, and the failure mode and confidence probability are output. The enhancement module dynamically adjusts the tower risk level based on the failure mode and confidence probability, generates monitoring strategy instructions, increases the sampling frequency of specific channels, and pre-allocates computing resources to special analysis algorithms. The correlation module performs real-time spectrum analysis on the improved sampling frequency, identifies specific characteristic spectra, re-analyzes the surface deformation sequence in the multi-source standardized dataset, detects nonlinear trends, and triggers special early warnings during spatiotemporal correlation matching. The fusion module integrates failure modes and confidence probabilities with specific early warnings to generate a comprehensive diagnostic report.

[0014] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the tower foundation deformation monitoring method coupled SAR and optical fiber as described in the first aspect of the present invention.

[0015] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the method for monitoring the deformation of a tower foundation coupled with SAR and optical fiber as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By acquiring synthetic aperture radar images and fiber optic sensing data of the monitoring area, and performing temporal interferometry and spatiotemporal alignment standardization processing respectively, a multi-source standardized dataset is constructed. This dataset is input into a pre-trained dual-stream neural network model to extract macroscopic deformation spatial features and microscopic mechanical response features, which are then fused using an attention mechanism. The fused features are then weighted similarity matching calculations with a pre-set failure mode feature knowledge base to output specific failure modes and their confidence probabilities. Based on the failure modes and confidence probabilities, the tower risk level is dynamically adjusted, and a monitoring strategy is generated. The system instructs the user to increase the sampling frequency of a specific distributed acoustic sensing channel and pre-allocate computing resources to a specialized analysis algorithm. It then performs real-time spectral analysis on the data after increasing the sampling frequency to identify specific characteristic spectra. Simultaneously, it re-analyzes the surface deformation sequence to detect nonlinear trends. When the two are correlated and matched in time and space, a special early warning is triggered. By fusing failure modes, confidence probabilities, and special early warning information, a comprehensive diagnostic report is generated, including mechanistic diagnosis, risk level, and treatment recommendations. Through deep feature fusion and dynamic closed-loop feedback, it achieves early, highly reliable intelligent monitoring and early warning of tower foundation deformation, especially the risk of hidden erosion. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for monitoring tower foundation deformation by coupling SAR and optical fiber.

[0019] Figure 2 This is a schematic diagram of a tower foundation deformation monitoring system that couples SAR and optical fiber. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The appearance of an embodiment in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0023] Reference Figures 1-2 As one embodiment of the present invention, this embodiment provides a method for monitoring the deformation of a tower foundation coupled with SAR and optical fiber, comprising the following steps: S1. Acquire SAR images and fiber optic sensing data of the monitoring area, perform temporal interferometry and spatiotemporal alignment standardization processing respectively, and construct a multi-source standardized dataset.

[0024] S1.1 Acquire SAR images and fiber optic sensing data of the monitoring area, perform time-series interferometry on the acquired SAR images of the monitoring area to obtain the surface deformation time series, deformation rate map and phase gradient field.

[0025] Furthermore, SAR images of the monitoring area are acquired by accessing the data stream from Sentinel-1 satellite covering the monitoring area. The acquired SAR images are processed using small baseline set temporal interferometry. This process first registers and resamples SAR images of all time phases based on satellite precise orbit data and external digital elevation model data. Then, interferometric image pairs are selected according to a set spatiotemporal baseline threshold, and a differential interferogram set is generated. The entangled phases in the differential interferogram set are unwrapped using a minimum cost flow algorithm based on Delaunay triangulation for two-dimensional phase unwrapping. This is combined with spatiotemporal filtering methods. Atmospheric delay phase estimation and removal are performed on the phase of the unwrapped differential interferogram using meteorological data from external global navigation satellite systems. Finally, singular value decomposition inversion is performed on the phase time series of all interferometric image pairs after atmospheric noise removal to solve for the displacement of each pixel in the radar line-of-sight direction over time, i.e., the surface deformation time series. At the same time, the deformation rate map is obtained by linear fitting of the surface deformation time series and calculating the rate of change in the time dimension. The spatial directional derivatives of the interferometric phase of each time phase in the east-west and north-south directions are obtained to obtain the phase gradient field that contains both magnitude and direction information and is used to characterize the spatial details of deformation.

[0026] Specifically, this method delves deeper into the results of temporal interferometry, exceeding conventional approaches. A key innovation is the proactive generation and retention of the phase gradient field as a core input for subsequent advanced analysis. Existing techniques typically culminate in surface deformation time series and deformation rate maps, with the phase gradient often treated as an intermediate product of the unwrapping process or noise to be filtered out. This method overturns this conventional understanding, recognizing that the phase gradient field is not simply noise, but a valuable source of information containing the spatial structural characteristics of the deformation field. The phase gradient is the first-order rate of change of deformation in space, directly related to the surface strain tensor, and is more sensitive to local deformation modes such as twisting, tilting, and uneven settlement than the deformation itself. For example, uniform settlement responds weakly to the phase gradient field, manifesting as large areas of low values; while uneven deformation caused by local instability of tower foundations (such as erosion or pile failure) forms significant spatial anomalies in the phase gradient field, such as high-gradient closed loops or directional stripes. By elevating the phase gradient field from an intermediate variable to a key output, this method provides subsequent steps with macroscopic observation data that is physically homologous to the strain field detected by fiber optic sensing (both reflect strain / deformation gradients) and can be directly compared and fused with high-dimensional features. This breaks the limitation of traditionally comparing only deformation-level data between space and ground, and lays an indispensable data foundation for realizing mechanism-level diagnosis based on deformation spatial distribution patterns.

[0027] S1.2 Perform spatiotemporal alignment and standardization on the acquired fiber optic sensing data of the monitoring area to obtain spatiotemporally aligned fiber optic sensing data.

[0028] Furthermore, acquiring fiber optic sensing data of the monitoring area is accomplished through a fiber Bragg grating sensor array deployed on the tower foundation structure. The fiber Bragg grating sensor array consists of multiple fiber Bragg grating sensors connected in series. The reflection center wavelength of each fiber Bragg grating sensor follows the Bragg reflection condition, and is related to the fiber grating period and the effective refractive index of the fiber core. Changes in external temperature and stress alter the fiber grating period and the effective refractive index of the fiber core, causing a shift in the reflection center wavelength. A definite linear relationship exists between the amount of reflection center wavelength shift and the temperature change and strain. Based on this linear relationship, the acquired reflection center wavelength data is calculated into temperature and strain value sequences corresponding to the sensor installation locations, with spatiotemporal alignment and calibration. The normalization process first utilizes high-precision global navigation satellite system measurement technology to obtain the accurate geodetic coordinates of each fiber Bragg grating sensor, and then transforms these coordinates to a geographic projection coordinate system consistent with the SAR imagery to achieve spatial alignment. Secondly, the time series of data acquired by the fiber Bragg grating sensors is synchronized with the absolute time of SAR imagery acquisition using linear interpolation to achieve time alignment. Polynomial fitting is then applied to the calculated temperature and strain value sequences to detrend them and eliminate long-term sensor drift. Finally, Z-fractional normalization is used to eliminate initial baseline differences between different fiber Bragg grating sensors. The result is spatiotemporally aligned fiber optic sensing data that is strictly aligned with the SAR data on a spatiotemporal reference and has undergone normalization processing. Specifically, this innovation manifests in the construction of a precise and reliable collaborative platform for heterogeneous data fusion. Its core innovation lies in a rigorous preprocessing procedure that transforms discrete, asynchronous, and dimensionlessly variable fiber optic sensor readings into standardized field data that is strictly aligned with SAR observations on a spatiotemporal grid and whose physical quantities are pure and comparable. Current technologies for processing fiber optic data often remain at the level of single-point monitoring and alarming, or simply perform basic network topology visualization. They lack the depth of preprocessing required for quantitative and mechanistic correlation with macroscopic remote sensing data. This innovation goes beyond simply cleaning and transforming the data; through a series of operations including coordinate mapping, time synchronization, detrending, and standardization, it essentially transforms the discrete fiber Bragg grating sensor network into a continuously distributed virtual sensing field within a unified spatiotemporal framework. Each pixel (i.e., spatiotemporal grid cell) in this virtual sensing field carries a physical quantity (such as strain or temperature) representing the local mechanical or thermal state at that location, obtained through interpolation or deduction from measured data. For example, stress redistribution in the soil beneath a tower foundation can lead to increased strain in specific areas. These changes are captured by a fiber optic sensor network and processed in this step, creating a spatially localizable strain anomaly region within a virtual strain field. This processing enables point information from the fiber optics, reflecting the internal mechanical state, to be directly compared and corresponded one-to-one with surface information from SAR, reflecting external geometric deformation, on the same spatiotemporal grid. This removes a fundamental technical obstacle to subsequent deep data analysis that moves from correlation of phenomena to mutual verification of mechanisms, and is a key preprocessing step in breaking down the barriers to fusion of space, air, and ground data.

[0029] S1.3. Construct a multi-source standardized dataset by integrating the time series of surface deformation, deformation rate map, phase gradient field and spatiotemporally aligned fiber optic sensing data.

[0030] Furthermore, the fusion process uses a regular geographic grid covering the monitoring area as a unified spatial reference and the time series acquired from SAR imagery as a unified temporal reference to construct a spatiotemporal index matrix. For each grid cell in the spatiotemporal index matrix, the deformation value of the entire time series corresponding to the center point of that grid cell is extracted from the surface deformation time series to form a deformation time series segment. The average deformation rate value of that grid cell is extracted from the deformation rate map. The phase gradient components of that grid cell in the east-west and north-south directions are extracted from the phase gradient field. From the spatiotemporally aligned fiber optic sensing data, the strain of all fiber Bragg grating sensors within a certain range around that grid cell is extracted. Using historical temperature data, the strain and temperature data of these discrete points are interpolated to the center of the grid cell using the Kriging spatial interpolation algorithm to obtain the strain field estimation sequence and temperature field estimation sequence of the grid cell. The deformation time series segment, average deformation rate value, east-west and north-south phase gradient component values, strain field estimation sequence and temperature field estimation sequence are spliced ​​and combined in a predetermined order to form a feature vector representing the comprehensive state of the spatiotemporal grid cell. The above operation is repeated for all grid cells in the spatiotemporal index matrix, and the feature vectors of all grid cells are organized according to spatial location and temporal order, thus forming a structured, multi-dimensional time series dataset, i.e., a multi-source standardized dataset.

[0031] Specifically, by utilizing spatiotemporal gridding and spatial interpolation techniques, data from different scales, sources, and physical meanings are proactively reshaped into a unified data volume that is fully aligned in spatiotemporal and feature dimensions. Existing technologies often employ passive data layer stacking, where results from different sources are displayed side-by-side on reports or maps after independent analysis processes, constructing a data cube. Each basic unit (a spatiotemporal grid) integrates comprehensive information from both space-based macroscopic observations and ground-based microscopic sensing. For example, for a tower location suspected of foundation erosion, its corresponding data cube unit not only includes potential settlement acceleration signals and deformation spatial distortion characteristics observed by InSAR, but also incorporates soil strain relaxation or foundation local stress anomalies sensed by surrounding fiber optic sensors through interpolation. This design eliminates the need for subsequent machine learning models (such as two-stream neural networks) to handle complex multi-source data alignment and correlation issues, allowing them to directly and in parallel read the pre-correlated macroscopic deformation and micromechanical features from each data unit. This fundamentally solves the key challenges faced by deep learning models when processing multi-source heterogeneous data, such as inconsistent inputs, mismatched feature scales, and spatiotemporal misalignment. It places complex data preprocessing work before feature learning, creating optimal conditions for the model to focus on learning high-level semantic features that are cross-modal and strongly correlated with failure modes.

[0032] S2. Input the multi-source standardized dataset into the pre-trained dual-stream neural network model, extract the macroscopic deformation space features and microscopic mechanical response features, and then fuse them.

[0033] S2.1 Extract the phase gradient field from the multi-source normalized dataset and input it into the macroscopic deformation feature extraction branch of the pre-trained two-stream neural network model; extract the high-density strain field from the multi-source normalized dataset and input it into the microscopic mechanical response feature extraction branch of the pre-trained two-stream neural network model.

[0034] Furthermore, from the multi-source normalized dataset, the phase gradient component values ​​of each grid cell in the east-west and north-south directions are extracted according to the spatiotemporal index order. The phase gradient component values ​​are organized into a two-channel two-dimensional feature map according to the grid space arrangement. This two-dimensional feature map is fed into the macroscopic deformation feature extraction branch of the pre-trained two-stream neural network model. Simultaneously, the statistical features of the strain field estimation sequence of each grid cell, such as the mean and standard deviation, are extracted from the same multi-source normalized dataset. The statistical feature values ​​are organized into a single-channel two-dimensional feature map according to the grid space arrangement. This two-dimensional feature map is fed into the micromechanical response feature extraction branch of the pre-trained two-stream neural network model. The macroscopic deformation feature extraction branch and the micromechanical response feature extraction branch of the pre-trained two-stream neural network model receive two-dimensional feature map inputs with different physical meanings, but the branch structure is composed of multi-layer convolutional neural networks, which are used to learn in parallel and extract deep feature expressions from their respective input data.

[0035] Specifically, parallel and independent deep feature learning channels were constructed for two observational data with vastly different physical properties and data structures, breaking down the barrier of feature confusion caused by early mixed input in traditional multi-source data fusion. Existing technologies often simply stitch together data from different sources into a single multi-channel image and input it into a single network, or process them separately and then perform late-stage decision fusion. It is recognized that the phase gradient field and the high-density strain field respectively carry two different dimensions of information: the spatial rate of change of geometric deformation and the local statistical distribution of mechanical state. There is a correlation between them, but they should not be forcibly mixed in shallow networks. For example, the two-dimensional feature map of the phase gradient field records the direction and density of deformation stripes, while the two-dimensional feature map of the strain field depicts the spatial location and intensity range of stress concentration. By designing a dual-stream architecture, the convolutional kernels of the macroscopic branch specifically learn to identify spatial pattern fingerprints left in the phase gradient field by foundation tilting, settlement, or uneven deformation, such as linear stripes or closed loops; at the same time, the convolutional kernels of the microscopic branch focus on learning to identify distribution pattern textures reflecting structural stress anomalies from the statistical features of the strain field, such as stress concentration areas or strain gradient bands. The parallel and customized feature extraction mechanism ensures that the two types of information from heaven and earth can fully explore the inherent patterns within their respective data during the process of upgrading to high-level semantic features.

[0036] S2.2. Based on the pre-trained dual-stream neural network model, the macroscopic deformation feature extraction branch processes the phase gradient field to obtain the macroscopic deformation spatial features.

[0037] Furthermore, the macroscopic deformation feature extraction branch of the pre-trained two-stream neural network model processes the two-dimensional feature map of the input phase gradient field. This branch contains multiple cascaded convolutional and pooling layers. The initial convolutional layer uses small-sized kernels to extract local features from the two-dimensional feature map of the phase gradient field, capturing basic spatial patterns such as edges, corners, and textures. Deeper convolutional layers expand the receptive field and combine low-level features to learn and abstract complex spatial structural patterns related to tower foundation failure, such as identifying the anisotropic distribution of the phase gradient field, the continuity or abrupt changes in gradient direction, and the geometry of high-gradient regions. Pooling layers downsample between convolutional layers to compress the spatial size of the feature map and enhance the translation invariance of the features. After multiple rounds of convolution and pooling operations, the output of the macroscopic deformation feature extraction branch is flattened by a fully connected layer and mapped to a fixed-length feature vector. This feature vector is the macroscopic deformation spatial feature, which encodes all high-level spatial pattern information related to potential failure mechanisms inherent in the input phase gradient field. Specifically, this method directly uses the phase gradient field, a specific physical quantity, as the target input of a deep learning model. The model is then trained to extract an abstract deformation pattern language for engineering failure diagnosis, rather than traditional deformation values. Existing InSAR-based technical analysis largely relies on deformation rate mapping and manual interpretation, or simple image processing to extract deformation regions. The breakthrough of this method lies in leveraging the powerful spatial feature learning capabilities of convolutional neural networks, enabling the model to automatically learn how deformation is organized and evolves in space from the phase gradient field. The phase gradient field is the first derivative of deformation with respect to space, and it is extremely sensitive to local distortions, rotations, and discontinuities in the deformation field. Through training, the macroscopic deformation feature extraction branch can internalize a series of spatial pattern prototypes corresponding to engineering failures. For example, for tilting failure caused by foundation eccentricity, the model can learn to identify systematic gradient direction patterns pointing towards the load direction in the phase gradient field; for erosion caused by local soil loss, the model can learn to identify closed-loop gradient anomalies forming around potential erosion points in the phase gradient field. The method of learning patterns directly from the gradient field bypasses the traditional path of complex spatial analysis and mechanical inversion required to derive deformation patterns from deformation variables, and provides a data-driven, end-to-end feature extraction paradigm. It enables the model to capture subtle but typical pattern-significant deformation spatial features that are difficult for the human eye to directly identify, enhancing the detection capability and pattern generalization ability of early signs of complex failure processes.

[0038] S2.3. Micromechanical response feature extraction based on pre-trained dual-stream neural network model: Branch processing of high-density strain field to obtain micromechanical response features.

[0039] Furthermore, the micromechanical response feature extraction branch of the pre-trained dual-stream neural network model processes the input high-density strain field two-dimensional feature map. The micromechanical response feature extraction branch is also composed of multiple cascaded convolutional and pooling layers. The initial convolutional layer scans the high-density strain field two-dimensional feature map to extract the basic patterns of local strain distribution, such as the distribution and boundaries of high and low strain value regions. Subsequent deep convolutional layers learn and abstract complex features that can characterize the overall or local mechanical state anomalies of the tower foundation by combining a wider range of local information. For example, they can identify the symmetry breaking of strain distribution, the morphology and evolution of strain concentration areas, and the interruption of the continuity of strain transmission paths. The pooling layer is used to reduce the feature dimension and filter noise to ensure that the extracted features are robust to small local changes. After a series of feature transformations, the output of the micromechanical response feature extraction branch is transformed into a feature vector with the same dimension as the output of the macroscopic deformation feature extraction branch through a fully connected layer. This feature vector is the micromechanical response feature, which condenses the key mechanical information contained in the high-density strain field that reflects the stress and strain state inside the structure.

[0040] Specifically, deep learning transforms these features into high-dimensional feature vectors representing the structural mechanical health state, achieving a leap from sensor reading sets to semantic abstraction of structural mechanics. Traditional fiber optic monitoring and analysis typically focus on over-threshold alarms from individual sensors or rely on expert experience to manually interpret strain contour maps. Training the micromechanical response feature extraction branch automatically makes it an expert in interpreting structural stress contour maps. Convolutional neural networks can learn mechanical patterns closely related to structural failure mechanisms from the spatial distribution of the strain field. For example, facing uneven foundation settlement, the model can learn to identify the distribution patterns of tensile-compressive strain pairs representing bending of the foundation slab in the strain field; facing frost heave, the model can learn to identify the widely distributed tensile strain region patterns at the bottom of the foundation; facing local foundation voids, the model can learn to identify abrupt changes or disappearances of strain support patterns near the support points. These micromechanical response features are no longer simple numerical values, but rather abstract semantics encoding the structural stress state, boundary condition changes, and potential damage locations. By extracting these features, an interpretable bridge based on mechanical mechanisms is provided for subsequent comparison and fusion with macroscopic deformation features. This allows the entire diagnostic process to go beyond just assessing how much the external appearance has changed, delving deeper into the internal stress state, significantly enhancing the physical reliability and mechanistic depth of the diagnostic conclusions, and breaking down the interpretive barriers between external monitoring and internal force perception.

[0041] S2.4. Use the attention mechanism to fuse macroscopic deformation space features and microscopic mechanical response features to obtain fused multimodal features.

[0042] Furthermore, the macroscopic deformation space feature vector output from the macroscopic deformation feature extraction branch and the microscopic mechanical response feature vector output from the microscopic mechanical response feature extraction branch are concatenated along the feature dimension to form an initial joint feature vector. This initial joint feature vector is then input into an attention weight generation network, which typically consists of fully connected layers and activation functions. This network is used to analyze the importance of each feature dimension in the initial joint feature vector for the failure mode discrimination of the current input sample. The attention weight generation network outputs an attention weight vector with the same dimension as the initial joint feature vector. Each scalar value in the attention weight vector corresponds to the importance weight of the same dimension feature in the initial joint feature vector. The attention weight vector and the initial joint feature vector are multiplied element-wise to achieve adaptive weighting enhancement or suppression of different feature dimensions. The weighted feature vector is then used as the fused multimodal feature. The fused multimodal feature integrates deformation space mode information from the phase gradient field and mechanical state information from the high-density strain field, and dynamically highlights the most discriminative part according to the current data context.

[0043] Specifically, an attention mechanism is employed to implement a data-driven, dynamically focused feature-level fusion strategy, fundamentally changing the static mode of fixed weights or simple splicing in traditional multi-source fusion. Existing technologies often assign fixed weights or treat features from different sources equally, failing to address the dynamic changes in the contribution of various features under complex scenarios as specific failure modes and data quality change. The attention mechanism is introduced as an intelligent feature arbitrator, allowing the model to automatically determine and emphasize the feature dimensions most relevant and consistent to the current identification task based on the specific macroscopic deformation spatial features and microscopic mechanical response features during inference, while weakening irrelevant or contradictory feature dimensions. For example, when macroscopic deformation features show a weak, suspected erosion-induced accelerated settlement spatial pattern, and microscopic strain features also detect an abnormal pattern of soil strain release in the same area, the attention mechanism will assign high weights to these mutually corroborating feature dimensions pointing to the same physical process (erosion), thereby strengthening this potential hazard signal in the fused multimodal features. Conversely, if macroscopic features show anomalies but microscopic features do not respond, or if the patterns indicated by the two are physically contradictory, the attention mechanism will reduce the weight of the corresponding features to suppress possible false alarms. This dynamic, content-aware fusion approach can effectively handle noise, missing information, or inconsistencies that may exist in terrestrial and terrestrial data due to differences in observation principles and spatiotemporal resolution. It achieves information complementarity and redundancy verification, ensuring that the generated fused multimodal features are the most discriminative and robust information integration. S3. Compare the fused macroscopic deformation space features and microscopic mechanical response features with the pre-set failure mode feature knowledge base, and output the failure mode and confidence probability.

[0044] S3.1. Based on the contribution weights of each dimension in the fused multimodal features to the failure mode discrimination, a feature dimension weight vector is formed.

[0045] Furthermore, this is achieved through an auxiliary neural network submodule that processes the fused multimodal features. This submodule receives the fused multimodal feature vector as input, passes it through one or more fully connected layers and nonlinear activation functions, and finally outputs a feature dimension weight vector with the same dimension as the fused multimodal feature vector. Each weight value in the feature dimension weight vector... The corresponding multimodal feature vector after fusion Features in each dimension Importance rating, weight value The sub-network analyzes and calculates the contextual information of the input features to reflect the discriminative power of the feature dimension in distinguishing different failure modes. The sub-network is trained together with the entire two-stream neural network model during the pre-training process. The parameters are optimized through the backpropagation algorithm so that the final feature dimension weight vector can dynamically and adaptively highlight the most critical discriminative feature dimension in the current input scenario, while suppressing the influence of noise or irrelevant feature dimensions.

[0046] Specifically, it is recognized that although the fused multimodal features integrate information from both the ground and space, the contribution of different feature dimensions to identifying the specific failure mode is not equal and varies depending on the specific type of hazard reflected by the input data. For example, when the monitoring scenario points to a foundation tilting pattern, feature dimensions that encode the consistency of the deformation direction gradient and the strain concentration on one side of the foundation and the strain release on the other side become crucial; while when the scenario points to frost heave, feature dimensions reflecting temperature changes and the tensile strain distribution at the bottom of the foundation may become key discriminative criteria. By training a lightweight sub-network to dynamically generate feature dimension weight vectors, intelligent weighting tailored to specific situations is achieved. This sub-network acts as a feature attention filter, automatically focusing on the most discriminative feature subset based on the global context of the currently fused features.

[0047] S3.2. Use the feature dimension weight vector to perform weighted similarity matching calculation on each failure mode feature in the fused multimodal features and the pre-set failure mode feature knowledge base to obtain a set of weighted similarity values.

[0048] Furthermore, a weighted similarity matching calculation is performed on the fused multimodal features and each failure mode feature in the pre-set failure mode feature knowledge base using the feature dimension weight vector, resulting in a weighted similarity value set. The calculation process iterates through each failure mode feature vector in the pre-set failure mode feature knowledge base. For the first... Each failure mode feature vector is used to calculate a weighted similarity value between itself and the fused multimodal feature vector, according to a weighted similarity expression. The calculation yields the corresponding value for each failure mode feature in the pre-set failure mode feature knowledge base. After that, all The set of values ​​constitutes a weighted similarity value set, which quantifies the degree of directional consistency between the fused multimodal features and each known failure mode feature in the knowledge base in the weighted feature space.

[0049] Specifically, the feature dimension weight vector is used as a spatial transformation operator in similarity calculation. This operator performs anisotropic scaling transformation on the feature space, and when calculating the dot product and magnitude, it scales the weights of important feature dimensions (high weights). Amplify the secondary or interfering dimensions (low weights) The similarity matching process involves contraction. For example, when matching local erosion patterns, if the feature dimension weight vector indicates that certain feature dimensions reflecting the correlation of local high-frequency vibration spectrum and the acceleration of small deformations have high weights, then even if the feature to be identified and the erosion template feature have only a slight positive consistency in these dimensions, they will be significantly amplified and contribute to the final similarity score. Conversely, in other dimensions with low weights, even if there are large differences, the impact on the final score is small. Weighted similarity matching is essentially a measurement performed in a discriminative subspace that is adaptively constructed based on the current data and is most conducive to differentiation. It makes the matching process insensitive to noise and irrelevant feature changes, while being highly sensitive to true pattern features, thereby improving the accuracy and robustness of pattern recognition. Its advantages are particularly evident in multi-source fusion scenarios with high feature dimensions, information redundancy, and complex noise.

[0050] The weighted similarity expression is: ; in, To integrate the fused multimodal features with the pre-defined failure mode feature knowledge base... The weighted similarity value between each failure mode feature This represents the total number of feature dimensions for the fused multimodal features and each failure mode feature. The feature dimension weight vector at the th Weight values ​​on each feature dimension For the first in the pre-set failure mode feature knowledge base The failure mode feature vector at the th failure mode feature vector in the th ... Feature values ​​in each feature dimension This refers to the index of the failure mode features in the pre-defined failure mode feature knowledge base. This is the dimension index of the feature vector. For the first The original eigenvalues ​​of each eigenvector.

[0051] S3.3 Select the failure mode feature corresponding to the maximum value from the weighted similarity value set as the matching failure mode feature, and map the matching failure mode feature to the failure mode.

[0052] Furthermore, the failure mode feature corresponding to the maximum value in the weighted similarity value set is selected as the matching failure mode feature, and the matching failure mode feature is mapped to a failure mode. The operation first scans the weighted similarity value set, finds the weighted similarity value with the largest value in the set, and retrieves the failure mode feature vector from the pre-set failure mode feature knowledge base based on the index in the weighted similarity value set. This retrieved failure mode feature vector is determined as the matching failure mode feature. Then, by querying the mode label mapping table associated with the pre-set failure mode feature knowledge base, the matching failure mode feature is mapped to a specific and readable failure mode name, such as uniform settlement, tilting, local erosion, or frost heave. This mapping process establishes the association from the abstract high-dimensional feature vector to the failure mode category of specific engineering semantics, thereby completing the determination of the most likely failure mode to which the current monitoring state belongs.

[0053] Specifically, the selection of the maximum value is based on a higher-quality decision space optimized by intelligent feature weighting and adaptive metric. Since the weighted similarity value set has highlighted key discriminative features and suppressed interference through feature dimension weight vectors, the maximum value more accurately represents the true and essential matching degree between the feature to be identified and a pattern in the knowledge base. For example, without weighting, a normal state feature affected by noise may have moderate but similar similarities with multiple failure mode features, resulting in a less prominent maximum value and ambiguous decision. After weighting, key feature dimensions related to real hidden dangers (such as erosion) are amplified, making the similarity corresponding to the erosion pattern potentially higher than other patterns, thus making the maximum value more prominent and the decision more certain. This two-stage design—optimizing the metric first and then selecting the decision—ensures that the final mapped failure mode is based on the most reliable and focused similarity evidence, significantly improving the accuracy of pattern discrimination and the confidence of the decision, while reducing the risk of misjudgment and missed judgment.

[0054] S3.4 Convert the maximum value in the weighted similarity set into a confidence probability using a normalization function, and output the failure mode and confidence probability.

[0055] Furthermore, the maximum value in the weighted similarity set is converted into a confidence probability using a normalization function, outputting the failure mode and confidence probability. Then, the entire weighted similarity set is processed using the softmax normalization function, which applies the maximum value to each weighted similarity value. Perform exponentiation on the value and divide by all The value exponent and sum, thus making each The value is converted into a probability value between zero and one, and the sum of all converted values ​​is one. The probability value obtained after the maximum value is transformed by softmax is used as the confidence probability of the output failure mode. This confidence probability, together with the failure mode name, constitutes the output. The confidence probability quantifies the relative certainty of the failure mode. For example, a high confidence probability indicates a high degree of consistency between the fused multimodal features and the matched failure mode features in the weighted discriminant space, while a low confidence probability suggests that the matching result may be uncertain or that the current state is significantly different from all modes in the knowledge base.

[0056] Specifically, the weighted similarity set is normalized using the softmax function, transforming it into a probability distribution. This probability value (confidence probability) has profound implications: it reflects not only the absolute similarity between the feature to be identified and the best matching pattern, but also the relative advantage of the best matching pattern compared to all other candidate patterns. For example, even if the absolute value is not high, if the similarity of all other patterns is much lower, the best matching pattern may still obtain a high confidence probability after softmax transformation. This indicates that although the current state is not typical, it is relatively the closest to that pattern. Conversely, if the similarity with the second and third highest similarity patterns is lower, the confidence probability may be lower. If the values ​​are very similar, the confidence probability will be low, indicating ambiguity in the decision and requiring further verification. The way failure modes and confidence probabilities are output provides crucial gradient information for subsequent dynamic risk adjustments, monitoring strategy optimization, and manual review, rather than simply a binary yes / no judgment. This makes the entire intelligent diagnostic process more interpretable and valuable for decision support, allowing operations personnel to take different levels of urgency based on the confidence probability, achieving an intelligent advancement from hard classification to soft decision-making.

[0057] S4. Dynamically adjust the tower risk level based on failure mode and confidence probability, generate monitoring strategy instructions, increase the sampling frequency of specific channels, and pre-allocate computing resources for special analysis algorithms.

[0058] S4.1 The failure mode and confidence probability are input into the preset risk level assessment rules to obtain the updated tower risk level.

[0059] Furthermore, the failure modes and confidence probabilities are input into preset risk level assessment rules, which exist in the form of tables or logic trees. These rules define the risk levels mapped to different failure modes and their corresponding confidence probability intervals. For example, local erosion failure mode combined with high confidence probability is mapped to a high risk level, uniform settlement failure mode combined with medium confidence probability is mapped to a medium risk level, and any failure mode combined with low confidence probability is mapped to a continuous observation level. When a specific set of failure modes and confidence probabilities is input, the preset risk level assessment rules determine the unique risk level corresponding to the combination through a lookup matching logic, thereby outputting the updated tower risk level. The updated tower risk level reflects the urgency assessment of the quantified safety hazards based on the current intelligent diagnostic results.

[0060] Specifically, an intelligent decision-making bridge was constructed that combines qualitative mechanism interpretation with quantitative confidence measurement and maps it to a tiered action guideline, surpassing traditional risk assessments based on single thresholds or simple logic. Existing risk assessment technologies often rely on single physical quantity thresholds such as whether deformation exceeds limits, lacking consideration of failure mechanisms and uncertainty measurement of judgment results. The core input of its preset risk level assessment rules is the failure mode and confidence probability output after deep feature fusion and intelligent matching. This means that the risk level depends not only on what happened (failure mode) but also on how confident the judgment is (confidence probability). For example, even if a high-risk mode like localized erosion is identified, if the confidence probability is low, the rule may assess it as medium risk or require verification, rather than directly triggering the highest alarm. This effectively avoids resource misallocation and false alarms caused by model misjudgment. Conversely, for uniform settlement with high confidence probability (usually lower risk), the rule may still assign medium risk to attract attention. Rules based on the two-dimensional input of mode and probability achieve an upgrade from phenomenon-based alarms to intelligent judgment-based alarms. It makes risk warnings more refined and intelligent, and can dynamically adjust the warning level based on the quality of the diagnostic conclusions and the nature of the hidden dangers, providing a scientific and reliable basis for subsequent differentiated and precise resource control measures.

[0061] S4.2 Generate monitoring strategy instructions based on the updated tower risk level; execute the monitoring strategy instructions to increase the sampling frequency of the distributed acoustic wave sensor (DAS) channel related to failure modes in the fiber optic sensor network monitoring data to a preset value.

[0062] Furthermore, monitoring strategy instructions are generated based on the updated tower risk level. These instructions include specific adjustment commands for the data acquisition parameters of the fiber optic sensor network. When executing these instructions, the key tower foundation areas or structural parts requiring monitoring are determined based on the updated tower risk level and the identified specific failure modes. This allows the location of the corresponding Distributed Acoustic Sensing (DAS) channels in the fiber optic sensor network monitoring these areas or parts. Then, instructions are sent to the fiber optic modems controlling these DAS channels to increase the data acquisition sampling frequency of the specified DAS channels from the conventional monitoring frequency to a higher preset value. The preset value is pre-set according to the requirements of monitoring accuracy and real-time performance for different risk levels and failure modes. For example, for the high-risk local erosion mode, the preset value is set to a frequency close to the highest sampling capability of the fiber optic modem to capture richer high-frequency vibration details, thereby obtaining DAS data with an increased sampling frequency.

[0063] Specifically, it achieves a leap in the adaptive capability of the monitoring system from uniform defense to focused monitoring, the core of which is the dynamic scheduling of on-demand sensing resources based on risk diagnosis results. Existing monitoring systems typically have fixed or periodically adjusted sampling frequencies, failing to provide immediate and accurate responses to sudden, specific risks. This system directly and automatically links the configuration of monitoring resources (specifically the sampling frequency of the distributed acoustic wave sensor (DAS) channels) with the updated tower risk levels and failure modes output by the front-end intelligent diagnostics. It establishes a closed loop of diagnosis-driven resource optimization. When a high-risk specific pattern (such as localized erosion) is identified, the monitoring strategy does not indiscriminately increase the frequency of all sensors, but precisely increases the frequency of only the specific DAS channels most relevant to the potential hazard location. For example, if the diagnosis indicates a risk of erosion at the southeast corner of the foundation, only the sampling rate of the DAS channels deployed near the southeast corner is increased. Precise control that pinpoints specific areas, with minimal resource expenditure (upgrading only a few key channels), acquires the most relevant and detailed data (high-frequency vibration signals) for potential hazards, providing a high-quality data foundation for subsequent specialized analysis. Secondly, it avoids the enormous data storage and transmission pressure brought about by high-frequency sampling across the entire network. It reflects the intelligence level of the monitoring system, which, like an experienced inspector, can immediately approach and carefully listen to suspicious signs, thereby significantly improving the early detection capability and monitoring efficiency of hidden and sudden hazards.

[0064] S4.3 Execute the monitoring strategy instructions and pre-allocate computing resources for the specialized analysis algorithm that identifies the erosion characteristic spectrum.

[0065] Furthermore, the monitoring strategy instruction is executed to pre-allocate computing resources for the specialized analysis algorithm for identifying erosion feature spectrum. After parsing the updated tower risk level and failure mode, if it is determined that a specialized analysis for a specific hidden danger (such as erosion) needs to be initiated, a computing resource pre-allocation instruction is generated. This instruction is sent to the computing unit responsible for running the specialized analysis algorithm. The computing unit reserves the necessary number of processor cores and memory space for the specialized analysis algorithm for identifying erosion feature spectrum, ensuring that the specialized analysis algorithm can be immediately scheduled and executed after acquiring the distributed acoustic wave sensor (DAS) data with the increased sampling frequency, without waiting for the resource application and allocation process. The amount of pre-allocated computing resources is pre-set according to the complexity and real-time requirements of the specialized analysis algorithm for identifying erosion feature spectrum. For example, dedicated computing cores and sufficient cache are allocated to the algorithm threads that run real-time spectrum analysis, feature extraction, and pattern matching.

[0066] Specifically, by shifting the preparation of computing resources from a passive response to proactive preparation, zero-wait startup of the data analysis phase is achieved. This perfectly synergizes with the on-demand sensing of the data acquisition phase, together forming a forward-looking resource guarantee mechanism for intelligent monitoring. In existing technical processes, computing tasks typically only request and schedule resources after data arrives. During sudden high-load analysis tasks, resource contention may cause processing delays, missing the optimal analysis opportunity. Simultaneously triggering the pre-allocation of computing resources while generating instructions to increase the sampling frequency is a forward-looking decision: since the decision has been made to collect more refined data (increasing the sampling frequency) to investigate specific risks (such as erosion), the computing resources required for the specialized analysis algorithms processing this high-refinement data should also be prepared in advance. For example, when the monitoring strategy requires an increase in the sampling rate of a specific distributed acoustic wave sensor (DAS) channel due to a high-risk erosion warning, it also instructs the computing unit to reserve sufficient CPU cores and memory for subsequent complex real-time spectrum analysis algorithms. This ensures that high-value data (high-frequency vibration data) can be analyzed immediately and without delay once it is generated, greatly shortening the end-to-end delay from data acquisition to warning decision-making. It incorporates computing resource management into a risk-based dynamic decision-making closed loop, enabling the entire system to make a rapid and coherent response when facing sudden risks. This truly realizes the real-time and intelligent nature of the monitoring-analysis-decision chain and is one of the key enabling links to ensure that the entire method can achieve very early warning.

[0067] S5. Perform real-time spectrum analysis on the improved sampling frequency, identify specific characteristic spectra, re-analyze the surface deformation sequence in the multi-source standardized dataset, detect nonlinear trends, and trigger special early warnings when matching spatiotemporal correlations.

[0068] S5.1 Perform real-time spectrum analysis on the distributed acoustic wave sensor (DAS) data after increasing the sampling frequency to obtain the real-time spectrum analysis results.

[0069] Furthermore, real-time spectrum analysis was performed on the DAS data after the sampling frequency was increased. The real-time spectrum analysis adopted the short-time Fourier transform method, which divided the continuous DAS data stream after the sampling frequency was increased into an overlapping time window sequence. Windowing and Fourier transform were performed on the data in each time window, and the energy distribution of the signal in each frequency component in the time window was calculated to obtain a series of spectrum diagrams that evolved over time. These spectrum diagrams constituted the real-time spectrum analysis results. The real-time spectrum analysis results dynamically characterized the characteristics of the frequency components of the vibration signal in the DAS data after the sampling frequency was increased as a function of time.

[0070] Specifically, this method employs targeted high-resolution monitoring. Conventional monitoring, using a fixed sampling rate, may not be able to fully capture the high-frequency or weak vibration characteristics generated by certain potential hazards (such as erosion). Based on risk assessment, this method temporarily increases the sampling rate to a higher level when necessary, enabling real-time spectrum analysis to acquire a wider bandwidth and finer frequency resolution. For example, the acoustic emission signals generated by soil particles detaching in the early stages of erosion may have high frequencies and weak energy, which may be submerged or not fully recorded under conventional sampling rates. By increasing the sampling frequency, real-time spectrum analysis can observe these weak but crucial signal components at higher frequency bands. This strategy of providing analytical data on demand and then performing high-precision analysis ensures that subsequent feature identification steps obtain sufficient and high-quality data input, which is a key prerequisite for effectively detecting concealed, high-frequency potential hazards.

[0071] S5.2 Use real-time spectrum analysis results to identify specific characteristic spectra and obtain the identified specific characteristic spectra.

[0072] Furthermore, specific characteristic spectra are identified using real-time spectrum analysis results. The identified specific characteristic spectra are obtained by comparing the spectrum obtained from the real-time spectrum analysis results with a pre-built erosion characteristic spectrum library using a method based on pattern matching or machine learning classifiers. The erosion characteristic spectrum library stores typical vibration spectrum templates under different soil types and different erosion stages. The identification process calculates the similarity between the current spectrum and each template in the template library, or directly classifies the spectrum using a trained classifier. When the similarity exceeds a preset threshold or the classifier outputs a specific erosion category, it is determined that a specific characteristic spectrum has been identified in the current time period and at the corresponding sensor location. The specific characteristic spectra identified continuously or intermittently constitute the time series and spatial distribution information of the identified specific characteristic spectra.

[0073] Specifically, acoustic / vibration fingerprinting technology is introduced into the specific scenario of erosion in tower foundation hazards, and deeply integrated with prior real-time high-resolution spectrum analysis. Traditional methods either only monitor vibration amplitude or perform simple threshold judgments on the spectrum, lacking refined pattern recognition for specific physical processes (such as erosion). A pre-established erosion feature spectrum library is created, associating this complex physical process with observable vibration spectrum patterns. This is similar to creating a soundprint archive for the erosion process. When high-resolution spectrum data is input, the recognition algorithm searches for soundprints matching the archive within this data. For example, during erosion, sand grain rolling, collisions, and micro-fractures in the soil produce specific frequency combinations and energy distribution patterns, which are encoded in the feature spectrum library. Through pattern matching or classifiers, the algorithm can separate and identify these anomalous tones strongly correlated with erosion from the spectrum containing complex backgrounds such as environmental noise, wind vibration, and mechanical vibration. Fine-grained pattern recognition based on prior knowledge base enables the monitoring of erosion to no longer rely on the general magnitude of vibration, but on its unique acoustic characteristics, which greatly improves the specificity and accuracy of identification and reduces false alarms caused by other vibration sources.

[0074] S5.3. Reanalyze the surface deformation sequence in the multi-source standardized dataset to obtain the surface deformation sequence reanalysis results.

[0075] Furthermore, the surface deformation sequence in the multi-source standardized dataset is reanalyzed to obtain the surface deformation sequence reanalysis results. The surface deformation time series data corresponding to the currently concerned tower and surrounding area are extracted from the multi-source standardized dataset. This surface deformation time series is the result of previous time series interferometry. The reanalysis focuses on the latest part of the sequence or the segment starting from a specific time point. The time series decomposition method is used to decompose the surface deformation time series into trend term, period term and residual term. Special attention is paid to the changes in the trend term, and higher-order statistical features of the sequence such as autocorrelation function or skewness may be calculated to reveal its dynamic characteristics. The surface deformation sequence reanalysis results are a feature set that is more reflective of its recent change patterns and potential anomalies after in-depth mining of the original deformation sequence.

[0076] Specifically, when failure mode identification (FMID) indicates a specific risk (such as scouring), it triggers a targeted review of deformation data at the same location. This review goes beyond simply observing deformation; it utilizes more sensitive analytical tools (such as time series decomposition and higher-order statistics) to detect potential nonlinear precursors in the series that characterize accelerated deformation or instability. For example, a series that initially showed slow, linear settlement may begin to exhibit a slight increase in settlement rate or volatility under scouring. These signs might be masked by the long-term linear trend, but can be detected through targeted reanalysis (such as observing the expansion of residual terms or subtle changes in the slope of trend terms). This embodies the idea of ​​risk-guided analysis, allowing data analysis resources to focus on the most likely areas of problem and the most critical signal characteristics, improving the efficiency and sensitivity of extracting early warning information from deformation data.

[0077] S5.4 Detect nonlinear trends from the reanalysis results of the surface deformation sequence and obtain the detected nonlinear trends.

[0078] Furthermore, nonlinear trends are detected from the reanalysis results of the surface deformation sequence. The detection of nonlinear trends mainly targets the trend term in the reanalysis results of the surface deformation sequence or the smoothed sequence. Statistical tests such as the Mann-Kendall trend test are used to analyze whether the sequence has a monotonically increasing or decreasing trend. Piecewise linear fitting or nonlinear regression models (such as exponential and logarithmic models) are further used to quantify whether the rate of change of the trend changes over time, and to determine whether the deformation process changes from a linear or static state to an accelerating or decelerating state. The detection result is a Boolean sign or a quantitative index used to indicate whether the surface deformation sequence has shown a statistically significant nonlinear accelerating or decelerating trend within the current analysis time window, i.e., the detected nonlinear trend.

[0079] Specifically, the focus is on the dynamic changes in the deformation process. For progressive damage such as erosion, the rate of settlement or tilting may have already undergone detectable changes before measurable and significant displacement of the foundation occurs. Detecting nonlinear trends aims to capture signs of acceleration. For example, the Mann-Kendall test can be used to determine whether the settlement rate has recently shown a statistically significant increasing trend; nonlinear fitting can quantify the degree of acceleration. The strategy of focusing on changes in the rate of change allows early warnings to be triggered much earlier than the time when the deformation itself reaches a dangerous threshold, achieving true trend warning or early warning. It shifts the focus of early warning from how much damage has already occurred to how quickly it is deteriorating, which is crucial for taking preventative maintenance measures and avoiding catastrophic consequences.

[0080] S5.5 When the identified specific feature spectrum matches the detected nonlinear trend in spatiotemporal correlation, a special early warning is triggered.

[0081] Furthermore, when the identified specific feature spectrum matches the detected nonlinear trend in a spatiotemporal correlation, a special warning is triggered. The spatiotemporal correlation matching first requires that the geographic location of the sensor corresponding to the identified specific feature spectrum and the location of the surface deformation pixel corresponding to the detected nonlinear trend overlap spatially or are within a preset proximity range. Secondly, it requires that the time period of the identified specific feature spectrum and the time window corresponding to the detected nonlinear trend are synchronized in time or have a preset sequential correlation logic. For example, the vibration spectrum feature appears first or appears simultaneously with the deformation acceleration trend. When both the spatial location correlation and temporal synchronization conditions are met, the spatiotemporal correlation matching is determined to be successful. Once the matching is successful, a special warning for the specific hidden danger (such as erosion) is triggered. The special warning includes the specific hidden danger type, the location of occurrence, the correlation evidence (spectral map and deformation trend curve), and the warning level determined according to the risk level.

[0082] Specifically, the warning system requires two pieces of evidence from completely different observation principles (acoustic vibration and geometric deformation) that point to the same physical process (such as erosion). For example, detecting only abnormal vibration (which may be due to construction interference) or only detecting only slight acceleration of deformation (which may be caused by temperature changes) will not trigger an erosion warning. Only when, at the same location and time, both the characteristic sound of erosion is heard (identified specific characteristic spectrum) and signs of accelerated foundation subsidence are observed (detected nonlinear trend) will the alarm be triggered. This two-factor correlation verification mechanism utilizes the independence and complementarity between different physical observations: the vibration signal provides the specificity of process identification (what is happening), and the deformation trend provides direct evidence of the mechanical consequences (what results were caused). The spatiotemporal coupling of these two factors constitutes a robust and reliable evidence loop, revolutionizing the credibility of the warning (i.e., reducing the false alarm rate). Simultaneously, it improves the targeting (clearly identifying erosion rather than other problems) and early warning (providing a warning based on initial vibration characteristics before significant acceleration of deformation), which is the fundamental guarantee for achieving highly reliable and accurate intelligent warnings.

[0083] S6 integrates failure modes and confidence probabilities with specific early warnings to generate a comprehensive diagnostic report.

[0084] S6.1. Correlate and align the failure modes and confidence probabilities with the special early warning in the spatiotemporal dimension to obtain multi-source diagnostic information after correlation and alignment.

[0085] Furthermore, failure modes and confidence probabilities are correlated and aligned with special early warnings in the spatiotemporal dimension. Failure modes and confidence probabilities include one or more potential failure modes and their corresponding confidence probabilities output by a dual-stream neural network model. These modes and confidence probabilities are associated with specific tower identifiers or geographical areas. Special early warnings include alarms triggered by multi-physical quantity evidence chains for specific hazard types, such as erosion. The alarms are associated with specific sensor locations and early warning timestamps. The correlation and alignment operation first uses a geographic information system to spatially match the tower or area locations indicated by the failure modes and confidence probabilities with the sensor locations indicated by the special early warnings to confirm whether they point to the same or adjacent monitoring targets. Secondly, the time of failure mode identification is compared with the time of special early warning triggering to ensure that the two are correlated in the time window. For example, failure mode identification occurs first and the early warning occurs after a reasonable delay. Finally, the successfully matched failure modes and confidence probabilities are bound with the special early warning information to form a set of multi-source diagnostic information that is interconnected and mutually corroborated in spatial location and temporal logic, i.e., the multi-source diagnostic information after correlation and alignment.

[0086] Specifically, it mandates strict spatiotemporal alignment of failure modes and confidence probabilities with specific early warnings, based on a spatiotemporal logic-driven evidence verification and integration process. For example, a dual-stream neural network model might output a failure mode and confidence probability of 70% for basic local erosion, while a specific early warning triggers a warning for erosion at the southeast corner of the foundation. The alignment operation verifies: whether the local erosion indicated by the failure mode matches the erosion type warned of; whether the tower location identified by the failure mode spatially overlaps with the sensor location warned of; and whether the time of failure mode identification is earlier than or synchronized with the warning time. Only when all spatiotemporal and logical conditions match is the information bound together. This process ensures that the final output diagnostic information is not isolated, contradictory fragments, but a complete chain of evidence that supports each other in time, space, and physical mechanisms. It automatically completes the cross-validation work that traditionally required manual intervention, improving the internal consistency and reliability of diagnostic conclusions.

[0087] S6.2 Integrate and align the multi-source diagnostic information according to the preset report template to generate a draft diagnostic report.

[0088] Furthermore, the multi-source diagnostic information, after being integrated and aligned according to a preset report template, is used to generate a draft diagnostic report. The preset report template is a structured document framework containing several fixed fields such as title, time, monitoring target location, mechanistic diagnostic conclusion, risk level assessment, key evidence display, and preliminary treatment recommendations. During the integration process, the contents of the multi-source diagnostic information after being integrated and aligned are filled into the corresponding fields of the report template. For example, the names and confidence probabilities of the main failure modes in the failure modes and confidence probabilities are filled into the mechanistic diagnostic conclusion field; the risk level assessment rules, based on the failure modes, confidence probabilities, and special early warnings, are filled into the risk level assessment field; and key data or chart links or thumbnails, such as spectrum diagrams, deformation trend curves, and failure mode matching similarity, which serve as evidence in the multi-source diagnostic information after being integrated and aligned, are filled into the key evidence display field. Based on the filled content, a structured draft diagnostic report containing all key information is automatically generated.

[0089] Specifically, an intelligent report generator was defined. This report template is not merely an empty shell, but rather embeds information organization logic: it knows how mechanistic diagnostic conclusions, risk levels, and evidence should be arranged and presented to support decision-making. The integration process essentially involves automatically disassembling and filling the verified evidence package of multi-source diagnostic information—after correlation and alignment—into the corresponding positions in the template. For example, it automatically summarizes erosion, 70% confidence probability, and high risk into mechanistic diagnostic conclusions and risk levels, and automatically correlates the corresponding spectrograms and deformation acceleration curves as key evidence. This process replaces tedious manual report writing, ensuring the uniformity of report format, the completeness of content, and the efficiency of generation. More importantly, it ensures that the output of each diagnosis follows the same logical framework, facilitating historical comparison, trend analysis, and standardized processing. This is not only an improvement in efficiency, but also a crucial step in transforming analytical results into standardized knowledge products that can directly support operational decisions.

[0090] S6.3 Format the draft diagnostic report and output a comprehensive diagnostic report including mechanistic diagnosis, risk level, and treatment recommendations.

[0091] Furthermore, the draft diagnostic report is formatted to output a comprehensive diagnostic report containing mechanistic diagnosis, risk level, and treatment recommendations. Formatting includes standardizing the text content in the draft diagnostic report, such as unifying fonts, font sizes, and paragraph formats; standardizing the graphical evidence in the draft diagnostic report, such as adding axis labels, units, legends, and necessary annotations to spectrum graphs and deformation trend curves, and ensuring that the graphs are clear and readable. At the same time, based on the content of the risk level assessment field and the preset treatment rule knowledge base, the draft diagnostic report is automatically matched and generated with corresponding preliminary treatment recommendation texts. For example, for high-risk erosion, it is recommended to immediately conduct on-site borehole verification and reinforcement preparation; for medium-risk uneven settlement, it is recommended to increase the monitoring frequency and arrange regular inspections. Finally, all the formatted text, graphs, and automatically generated treatment recommendations are integrated to output a comprehensive diagnostic report that is formatted in a standardized manner, rich in graphics and text, with clear conclusions and action guidelines.

[0092] Specifically, the deep integration of formatted output with intelligent decision support elevates the report from diagnostic report to decision-making recommendation. Traditional report formatting often focuses solely on aesthetics. The formatting process, however, embeds decision support logic. It goes beyond simply tidying the layout; it automatically generates preliminary, actionable recommendations based on the risk level identified in the draft report and a knowledge base of handling rules. For example, if the risk level field in the draft report is high-risk, the formatting program will retrieve preset recommendation templates for high-risk areas from the knowledge base, such as recommending a special inspection within 24 hours, using ground-penetrating radar or borehole equipment to conduct a detailed inspection of the warning area, and preparing necessary emergency reinforcement materials and plans. These recommendations, generated based on engineering experience and preset rules, are directly appended to the diagnostic conclusion. The resulting comprehensive diagnostic report is no longer just a technical document describing what happened and how serious it is, but an action guide that simultaneously provides recommendations on what to do. This extends the endpoint of data analysis to the starting point of operational decision-making, shortening the decision-making chain from perception to action and improving the response speed and practicality of the entire monitoring and early warning closed loop. This means that the monitoring system outputs proactive decision support, rather than passive information, truly enabling intelligent monitoring to empower intelligent operation and maintenance.

[0093] This embodiment also provides a tower foundation deformation monitoring system that couples SAR and optical fiber, including: a processing module that acquires SAR images and optical fiber sensing data of the monitoring area, performs temporal interferometry processing and spatiotemporal alignment standardization processing respectively, and constructs a multi-source standardized dataset; The comparison module inputs a multi-source standardized dataset into a pre-trained dual-stream neural network model, extracts macroscopic deformation space features and microscopic mechanical response features, and fuses them. The fused macroscopic deformation space features and microscopic mechanical response features are then compared with a pre-set failure mode feature knowledge base, and the failure mode and confidence probability are output. The enhancement module dynamically adjusts the tower risk level based on the failure mode and confidence probability, generates monitoring strategy instructions, increases the sampling frequency of specific channels, and pre-allocates computing resources to special analysis algorithms. The correlation module performs real-time spectrum analysis on the improved sampling frequency, identifies specific characteristic spectra, re-analyzes the surface deformation sequence in the multi-source standardized dataset, detects nonlinear trends, and triggers special early warnings during spatiotemporal correlation matching. The fusion module integrates failure modes and confidence probabilities with specific early warnings to generate a comprehensive diagnostic report.

[0094] This embodiment also provides a computer device applicable to the method for monitoring the deformation of tower foundations coupled with SAR and optical fiber, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the method for monitoring the deformation of tower foundations coupled with SAR and optical fiber as proposed in the above embodiment.

[0095] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0096] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the method for monitoring the deformation of a tower foundation coupled with SAR and optical fiber as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0097] In summary, this invention acquires synthetic aperture radar (SAR) images and fiber optic sensor data of the monitoring area, and performs temporal interferometry and spatiotemporal alignment standardization processing respectively to construct a multi-source standardized dataset. This dataset is then input into a pre-trained dual-stream neural network model to extract macroscopic deformation spatial features and microscopic mechanical response features, which are then fused using an attention mechanism. The fused features are then weighted similarity matching calculations with a pre-set failure mode feature knowledge base to output specific failure modes and their confidence probabilities. Based on the failure modes and confidence probabilities, the tower risk level is dynamically adjusted, and monitoring strategy instructions are generated. This system enhances the sampling frequency of specific distributed acoustic wave sensing channels and pre-allocates computing resources to specialized analysis algorithms. Real-time spectrum analysis is performed on the data after the sampling frequency is enhanced to identify specific characteristic spectra. Simultaneously, the surface deformation sequence is re-analyzed to detect nonlinear trends. When the two are correlated and matched in time and space, a specialized early warning is triggered. Failure modes, confidence probabilities, and specialized early warning information are integrated to generate a comprehensive diagnostic report with mechanistic diagnosis, risk level, and disposal recommendations. Through deep feature fusion and dynamic closed-loop feedback, early and highly reliable intelligent monitoring and early warning of tower foundation deformation, especially the risk of hidden erosion, is achieved.

[0098] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for monitoring the deformation of tower foundations coupled with SAR and optical fiber, characterized in that: This includes acquiring SAR images and fiber optic sensing data of the monitoring area, performing temporal interferometry and spatiotemporal alignment standardization processing respectively, and constructing a multi-source standardized dataset; The multi-source standardized dataset is input into a pre-trained dual-stream neural network model to extract macroscopic deformation space features and microscopic mechanical response features, and then fuse them. The fused macroscopic deformation space features and microscopic mechanical response features are compared with a pre-set failure mode feature knowledge base to output the failure mode and confidence probability. The risk level of the tower is dynamically adjusted based on the failure mode and confidence probability, monitoring strategy instructions are generated, the sampling frequency of specific channels is increased, and computing resources are pre-allocated to special analysis algorithms. Real-time spectrum analysis is performed on the improved sampling frequency to identify specific characteristic spectra. The surface deformation sequence in the multi-source standardized dataset is re-analyzed to detect nonlinear trends and trigger special early warnings when matching spatiotemporal correlations. By integrating failure modes and confidence probabilities with specific early warnings, a comprehensive diagnostic report is generated.

2. The method for monitoring tower foundation deformation by coupling SAR and optical fiber as described in claim 1, characterized in that: Acquire SAR images and fiber optic sensor data of the monitoring area, perform temporal interferometry and spatiotemporal alignment normalization processing respectively, and construct a multi-source normalized dataset, including the following steps: Acquire SAR images and fiber optic sensing data of the monitoring area, perform temporal interferometry on the acquired SAR images of the monitoring area, and obtain the surface deformation time series, deformation rate map and phase gradient field. Spatiotemporal alignment and standardization are performed on the acquired fiber optic sensing data of the monitoring area to obtain spatiotemporally aligned fiber optic sensing data. By integrating surface deformation time series, deformation rate map, phase gradient field and spatiotemporally aligned fiber optic sensing data, a multi-source standardized dataset is constructed.

3. The method for monitoring tower foundation deformation by coupling SAR and optical fiber as described in claim 2, characterized in that: The multi-source standardized dataset is input into a pre-trained two-stream neural network model to extract macroscopic deformation space features and microscopic mechanical response features, and then fused them, including the following steps: The macroscopic deformation feature extraction branch extracts the phase gradient field from the multi-source normalized dataset and inputs it into the pre-trained two-stream neural network model; the microscopic mechanical response feature extraction branch extracts the high-density strain field from the multi-source normalized dataset and inputs it into the pre-trained two-stream neural network model. The macroscopic deformation feature extraction branch based on the pre-trained two-stream neural network model processes the phase gradient field to obtain the macroscopic deformation spatial features; Micromechanical response feature extraction based on a pre-trained two-stream neural network model is used to process high-density strain fields and obtain micromechanical response features. The attention mechanism is used to fuse macroscopic deformation space features and microscopic mechanical response features to obtain fused multimodal features.

4. The method for monitoring tower foundation deformation by coupling SAR and optical fiber as described in claim 3, characterized in that: The fused macroscopic deformation space features and microscopic mechanical response features are compared with a pre-built failure mode feature knowledge base to output the failure mode and confidence probability, including the following steps: Based on the contribution weights of each dimension in the fused multimodal features to the failure mode discrimination, a feature dimension weight vector is formed. The feature dimension weight vector is used to perform weighted similarity matching calculation on each failure mode feature in the fused multimodal features and the pre-set failure mode feature knowledge base to obtain a set of weighted similarity values. Select the failure mode feature corresponding to the maximum value from the weighted similarity value set as the matching failure mode feature, and map the matching failure mode feature to the failure mode. The maximum value in the weighted similarity set is converted into a confidence probability using a normalization function, and the failure mode and confidence probability are output.

5. The method for monitoring tower foundation deformation by coupling SAR and optical fiber as described in claim 4, characterized in that: The tower risk level is dynamically adjusted based on failure mode and confidence probability, monitoring strategy instructions are generated, the sampling frequency of specific channels is increased, and computing resources are pre-allocated to specialized analysis algorithms, including the following steps: The failure mode and confidence probability are input into the preset risk level assessment rules to obtain the updated tower risk level; Based on the updated tower risk level, a monitoring strategy instruction is generated; the monitoring strategy instruction is executed to increase the sampling frequency of the distributed acoustic wave sensor (DAS) channel related to the failure mode in the fiber optic sensor network monitoring data to a preset value; The monitoring strategy instructions are executed, and specialized analysis algorithms that identify erosion characteristic spectrums are pre-allocated computing resources.

6. The method for monitoring tower foundation deformation by coupling SAR and optical fiber as described in claim 5, characterized in that: Real-time spectrum analysis is performed on the improved sampling frequency to identify specific characteristic spectra. The surface deformation sequences in the multi-source standardized dataset are re-analyzed to detect nonlinear trends. Specialized early warnings are triggered during spatiotemporal correlation matching, including the following steps: Real-time spectrum analysis was performed on the distributed acoustic wave sensor (DAS) data after the sampling frequency was increased, and the real-time spectrum analysis results were obtained. The specific characteristic spectrum is identified by using real-time spectrum analysis results. The surface deformation sequence in the multi-source standardized dataset is reanalyzed to obtain the surface deformation sequence reanalysis results; Nonlinear trends are detected from the reanalysis results of surface deformation sequences, and the detected nonlinear trends are obtained. A special early warning is triggered when the identified specific feature spectrum matches the detected nonlinear trend in a spatiotemporal correlation.

7. The method for monitoring tower foundation deformation by coupling SAR and optical fiber as described in claim 6, characterized in that: By integrating failure modes and confidence probabilities with specific early warnings, a comprehensive diagnostic report is generated, including the following steps: By correlating and aligning failure modes and confidence probabilities with specific early warnings in the spatiotemporal dimension, multi-source diagnostic information after correlation and alignment is obtained. The multi-source diagnostic information, after being integrated and aligned according to the preset report template, is used to generate a draft diagnostic report. The draft diagnostic report is formatted to output a comprehensive diagnostic report containing mechanistic diagnosis, risk level, and treatment recommendations.

8. A tower foundation deformation monitoring system coupled with SAR and optical fiber, based on the tower foundation deformation monitoring method coupled with SAR and optical fiber as described in any one of claims 1 to 7, characterized in that: This includes a processing module that acquires SAR images and fiber optic sensor data of the monitoring area, performs temporal interferometry and spatiotemporal alignment standardization processing respectively, and constructs a multi-source standardized dataset. The comparison module inputs a multi-source standardized dataset into a pre-trained dual-stream neural network model, extracts macroscopic deformation space features and microscopic mechanical response features, and fuses them. The fused macroscopic deformation space features and microscopic mechanical response features are then compared with a pre-set failure mode feature knowledge base, and the failure mode and confidence probability are output. The enhancement module dynamically adjusts the tower risk level based on the failure mode and confidence probability, generates monitoring strategy instructions, increases the sampling frequency of specific channels, and pre-allocates computing resources to special analysis algorithms. The correlation module performs real-time spectrum analysis on the improved sampling frequency, identifies specific characteristic spectra, re-analyzes the surface deformation sequence in the multi-source standardized dataset, detects nonlinear trends, and triggers special early warnings during spatiotemporal correlation matching. The fusion module integrates failure modes and confidence probabilities with specific early warnings to generate a comprehensive diagnostic report.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the tower foundation deformation monitoring method of any one of claims 1 to 7, which combines SAR and optical fiber.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the tower foundation deformation monitoring method of coupled SAR and optical fiber as described in any one of claims 1 to 7.

Images