Method for recognizing geological deformation through machine vision

Abstract

The invention discloses a method for recognizing geological deformation through machine vision. The method comprises the following steps of multi-source heterogeneous data periodic collection, multi-dimensional data preprocessing and feature enhancement, deformation field intelligent recognition driven by a mixed deep learning model, and dynamic threshold classification judgment of deformation types and risk levels. According to the method for recognizing geological deformation through machine vision, the problems of signal shielding and data blind areas of complex terrain are solved through multi-sensor optimization deployment, and maximum coverage of a monitoring area and continuous and complete data collection are achieved. Through multi-dimensional data preprocessing and multi-level fusion processing, data distortion is effectively corrected, noise is eliminated, weak features are highlighted, and the feature extraction identification degree and the classification result reliability are improved. The mixed deep learning model is coupled with the CNN and the LSTM, spatial structure features can be extracted and a dynamic process can be captured at the same time, and a high-resolution and high-precision geological deformation field map can be output.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geological monitoring and machine vision, and particularly relates to a method for identifying geological deformation by machine vision. BACKGROUND

[0002] In the field of geological deformation monitoring, traditional methods rely on single sensors, and the data sources are limited, making it difficult to obtain comprehensive and accurate geological deformation information. Although multi-source sensors can provide more abundant data, there are large differences in data characteristics of different sensors, and there are problems such as signal shielding and data blind area, and the data fusion and processing are difficult.

[0003] In terms of data processing, the traditional method has simple preprocessing operation, and it is difficult to effectively correct data distortion, eliminate noise interference, and highlight weak features. Feature extraction relies on artificial design, and the representation ability for complex geological textures and dynamic changes is limited.

[0004] In terms of geological deformation recognition model, the traditional model is difficult to capture spatial structure features and nonlinear dynamic process evolving over time, and lacks dynamic threshold self-adaptive adjustment mechanism for classification and determination of deformation type and risk level, which is low in accuracy and not timely in early warning.

[0005] In addition, the existing monitoring method has problems such as high processing delay, difficulty in meeting the needs of quasi-real-time monitoring, poor adaptability of hardware architecture to changes in model structure, and other problems, which restrict the development and application of geological deformation monitoring technology.

[0006] Therefore, it is necessary to provide a new method for identifying geological deformation by machine vision to solve the above technical problems. SUMMARY

[0007] To solve the above technical problems, the present application provides a method for identifying geological deformation by machine vision.

[0008] The method for identifying geological deformation by machine vision provided by the present application comprises the following steps:

[0009] S1: periodic acquisition of multi-source heterogeneous data: in the geological area to be monitored, a multi-sensor monitoring network composed of multiple types of sensors is deployed, and a predetermined acquisition period is set, the multi-sensor monitoring network at least includes an optical imaging device for obtaining high-resolution surface texture and morphology information, an interferometric synthetic aperture radar (InSAR) sensor for penetrating vegetation and clouds to obtain all-weather surface micro-deformation data, and a laser radar (LiDAR) sensor for generating high-precision three-dimensional digital elevation model (DEM); the step aims to periodically acquire multi-modal, multi-temporal remote sensing observation data sets covering the entire monitoring area.

[0010] S2: Multidimensional data preprocessing and feature enhancement step: A series of preprocessing operations are performed on the multi-modal, multi-temporal remote sensing observation data set collected in step S1, aiming to correct the data distortion introduced by sensors and environmental factors, eliminate noise interference, and highlight the weak features related to geological deformation. The preprocessing operations specifically include: radiometric calibration and atmospheric correction of optical images to eliminate the effects of light and atmosphere; high-precision geometric correction and orthorectification of all sensor data to eliminate geometric deformation caused by terrain undulations; and strict image registration of images from different periods and different sensors to ensure spatial alignment at the pixel level, providing a reference for subsequent change detection.

[0011] S3: Mixed deep learning model driven deformation field intelligent identification step: The time series data set preprocessed in step S2 is input into a pre-trained mixed deep learning model for geological deformation identification. The mixed deep learning model couples convolutional neural networks (CNN) and long short-term memory networks (LSTM). The convolutional neural network (CNN) module is responsible for extracting spatial structural features such as cracks, subsidence edges, and landslide body boundaries from single or differential images, generating feature maps representing deformation distribution. The long short-term memory network (LSTM) module is responsible for analyzing the time-ordered feature map sequence output by the CNN module to capture the nonlinear dynamic process and long-term dependence of geological deformation over time, and finally outputs a high-resolution, high-precision geological deformation field map covering the entire monitoring area.

[0012] S4: Dynamic threshold classification and determination step of deformation type and risk level: Based on the time series geological deformation field map generated in step S3, a dynamic threshold self-adaptive adjustment mechanism integrating state transition logic is used to finely divide the identified deformation area into different types of geological deformation and quantitatively evaluate the potential risk level. The dynamic threshold adjustment mechanism can automatically adjust the discrimination threshold for distinguishing different deformation states (e.g., stable, initial deformation, accelerated deformation, and unstable failure) based on real-time updated deformation field data and pre-set physical models, and combine with geological structure background knowledge base to finally classify the deformation area into specific types such as landslide, ground subsidence, and fault activity.

[0013] Preferably, the deployment scheme of the multi-sensor monitoring network in step S1 is optimized to solve the problem of signal obstruction and data blind area in complex mountainous terrain. The optimized deployment scheme includes: using an improved d-Xdraw visibility solving algorithm and three-dimensional visibility analysis to pre-calculate the effective coverage range of the sensor in undulating terrain, and combining a deployment solver based on an improved genetic algorithm or quantum particle swarm optimization algorithm to determine the optimal spatial layout position and attitude parameters of the sensor, thereby achieving maximum coverage of the monitoring area and continuity and integrity of data collection.

[0014] Preferably, the multi-dimensional data preprocessing and feature enhancement step in step S2 further comprises a multi-level data fusion process; the data fusion process is performed before the data is input into the hybrid deep learning model, and specifically comprises:

[0015] S21: Pixel-level fusion: fuse the registered multi-source images (such as optical images and InSAR intensity maps) at the pixel level to generate new images with richer texture and spectral information, thereby improving the recognition degree of subsequent feature extraction;

[0016] S22: Feature-level fusion: extract independent feature vectors from different sensor data (for example, extract texture features from optical images and deformation rate features from InSAR data), then splice or weighted combine these feature vectors to form a more discriminative multi-dimensional feature set;

[0017] S23: Decision-level fusion: use D-S evidence theory or Bayesian network method to comprehensively analyze and judge the preliminary classification results based on single data source, solve the possible conflicts and uncertainties between different data sources, and finally generate a more reliable and robust fusion decision result.

[0018] Preferably, before training, the convolutional neural network (CNN) module of the hybrid deep learning model in step S3 is pre-trained using a visual terrain classification method based on multi-encoding and feature fusion to enhance its representation ability for geological texture and geomorphic units; the pre-training process uses a large-scale geological image dataset, extracts SIFT features and performs PCA-Whitening preprocessing, combines GIST global features, constructs a multi-feature dictionary, and performs feature fusion through multi-kernel learning (MKL), trains a linear support vector machine (SVM) classifier, and finally initializes the weights of the SVM or the bottom convolution kernel of the CNN, thereby accelerating the convergence speed of the main task and improving the model generalization performance.

[0019] Preferably, the dynamic threshold self-adaptive adjustment mechanism in step S4 has a core of a mathematical framework based on state transition logic, which is used to accurately distinguish between landslide and ground subsidence and other types of geological deformation; the framework defines at least four geological states: stable state (S stable), initial creep state (S creep), accelerated deformation state (S accel), and failure state (S failure); the transition between states is driven by a transition function δ (S, I) → S', where the current state S and the input vector I jointly determine the next state S'; the input vector I includes quantitative parameters extracted from the deformation field map, such as maximum deformation rate (V max), cumulative displacement (D acc), deformation gradient, and external trigger factors (such as effective rainfall calculated based on the intensity-duration (I-D) curve model); the threshold values for state transition (for example, the rate threshold value for transitioning from S creep to S accel) are not fixed values, but are dynamically adjusted in real time by a dynamic function based on a four-parameter logistic regression model or Bayesian update, which can dynamically update the threshold values according to the statistical variability of historical monitoring data, thereby improving the accuracy of classification and the timeliness of early warning.

[0020] Preferably, the method is executed by a heterogeneous computing platform integrating a central processing unit (CPU) and a field programmable gate array (FPGA); wherein the control logic and complex decision-making involved in data acquisition (S1), data preprocessing (S2), and dynamic threshold classification (S4) are handled by the CPU, while the most computationally intensive deformation field intelligent identification step (S3) is completely offloaded to the FPGA for hardware acceleration to significantly reduce processing delay and meet the needs of quasi-real-time monitoring.

[0021] Preferably, the FPGA is designed and implemented with a deeply optimized hardware pipeline architecture to implement the convolutional neural network (CNN) calculation in step S3; the pipeline architecture decomposes complex convolution operations, pooling operations, and activation function calculations into multiple fine and continuous executable sub-stages, allowing multiple processing elements (PEs) to simultaneously calculate different regions of the input feature map through task parallelism and data parallelism, and further reducing the number of calculation steps and memory access times by fusing the linear transformation parameters of the BatchNorm layer with the convolution layer at the hardware level, thereby greatly improving data throughput and computational efficiency.

[0022] Preferably, the hardware pipeline architecture of the FPGA integrates an efficient memory access optimization strategy to minimize access delay to external DDR and other slow storage devices; the strategy includes:

[0023] S31: Data prefetching and weight caching: while the current convolution layer is being calculated, the convolution kernel weights and partial input feature map data required for the next layer calculation are prefetched from the off-chip DDR to the high-speed block random access memory (BRAM) on the FPGA chip for caching by a dedicated hardware prefetching logic;

[0024] S32: Double buffering ping-pong operation: two independent BRAMs are configured as double buffer areas for input feature maps and output feature maps respectively. When the calculation core is reading data from one buffer area for operation, the data transfer module (such as AXI-DMA controller) can write new data to the other buffer area at the same time, and vice versa. In this way, the data transfer delay is completely hidden, and seamless connection of calculation and data transfer is achieved, forming a "zero-copy" data flow of CNN feature map processing on the chip.

[0025] Preferably, in order to cope with the requirements of CNN model structure changes or multi-task parallel processing, a hardware task scheduling algorithm based on reconfigurable calculation is implemented on the FPGA; the scheduling algorithm (such as RDMS algorithm) can comprehensively evaluate the task correlation, resource utilization and internal communication overhead according to the CNN network topology structure, intelligently map and layout the calculation tasks with close communication (such as adjacent convolution layers) to the physically adjacent FPGA logic resources, and adjust the hardware function at runtime through dynamic partial reconfiguration (DPR) technology, thereby effectively reducing the data routing delay and unnecessary reconfiguration time in the FPGA, and further reducing the end-to-end processing delay.

[0026] Preferably, the CNN hardware interface circuit implemented on the FPGA has its module function definition and working timing realized through a global synchronous clock and a hierarchical synchronization mechanism to ensure the integrity and correctness of data in the "zero-copy" transmission process; the synchronization mechanism includes:

[0027] S311: interface layer synchronization: use the handshaking mechanism (such as VALID / READY signal) provided by the AXI bus protocol to ensure the synchronization of data transmission between the DMA controller and the BRAM buffer area;

[0028] S312: inter-module synchronization: within the FPGA, a central control module is designed to generate global synchronization clock signals and enable signals, which are distributed to the processing elements (PEs), cache controllers and data prefetching modules in the pipeline, to ensure that all modules trigger operations at the same clock edge;

[0029] S313: Cross-clock domain synchronization: For the different clock domains that may exist (for example, the DDR controller clock domain and the CNN computing core clock domain), an asynchronous FIFO (first-in-first-out queue) is used as a data buffer and synchronization bridge to safely transfer data and prevent the occurrence of metastability, thereby ensuring the stability and reliability of the timing of the entire hardware system.

[0030] Compared with the related art, the method for recognizing geological deformation by machine vision provided by the application has the following beneficial effects:

[0031] The method for recognizing geological deformation by machine vision provided by the application solves the problems of signal shielding and data blind area in complex terrain through optimized deployment of multiple sensors, and realizes maximized coverage and continuous and complete data acquisition of the monitoring area. Through multi-dimensional data preprocessing and multi-level fusion processing, data distortion is effectively corrected, noise is eliminated, weak features are highlighted, and the feature extraction recognition degree and the reliability of the classification result are improved. The hybrid deep learning model couples CNN and LSTM, can extract spatial structure features and capture dynamic processes at the same time, and outputs high-resolution and high-precision geological deformation field maps. The dynamic threshold self-adaptive adjustment mechanism accurately distinguishes deformation types and quantifies risk levels, improves classification accuracy and timeliness of early warning. The use of a heterogeneous computing platform and FPGA hardware acceleration significantly reduces processing delay and meets the needs of quasi-real-time monitoring. The deep optimization hardware pipeline architecture, memory access optimization strategy, reconfigurable computing scheduling algorithm and hardware synchronization mechanism of the FPGA improve the data throughput, computing efficiency and system stability, and enhance the adaptability of the hardware to changes in the model structure. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 The method for recognizing geological deformation by machine vision provided by the application is provided. DETAILED DESCRIPTION

[0033] The application will be further described below in conjunction with the drawings and embodiments.

[0034] In the specific implementation process, for example, Figure 1 A method for recognizing geological deformation by machine vision includes the following steps:

[0035] S1: Periodic acquisition of multi-source heterogeneous data step:

[0036] This step deploys a multi-sensor monitoring network composed of multiple types of sensors in the geological area to be monitored. The network includes:

[0037] Optical imaging device: used to obtain high-resolution surface texture and morphology information with a resolution better than 0.5 meters.

[0038] Interferometric Synthetic Aperture Radar (InSAR) sensor: used to penetrate vegetation and clouds to obtain all-weather ground surface micro-deformation data, with deformation measurement accuracy up to millimeter level.

[0039] LiDAR sensor: used to generate high-precision three-dimensional digital elevation model (DEM) with elevation accuracy better than 10 cm.

[0040] The sensor deployment scheme adopts an optimized d-Xdraw visibility solution algorithm for three-dimensional visibility analysis, and its mathematical model is as follows:

[0041] V(x,y,z)=∫∫∫ Ω f(θ,φ,ρ)·g(x,y,z)dθdφdρ

[0042] Where V(x,y,z) represents the visibility value at position (x,y,z), f(θ,φ,ρ) is the sensor observation model function, g(x,y,z) is the terrain occlusion function, and Ω is the sensor field of view angle range.

[0043] The optimal deployment position of the sensor is solved by an improved quantum particle swarm optimization algorithm (QPSO), and its position update formula is:

[0044] x i (t+1)=p i ±β·|m best -x i (t)|·ln(1 / u)

[0045] Where x i (t) represents the position of the i-th particle at time t, p i is the local optimal position, m best is the global optimal position, β is the contraction and expansion coefficient, and u is a random number uniformly distributed in the interval (0,1).

[0046] The deployment scheme of the multi-sensor monitoring network is optimized to solve the signal shielding and data blind area problems in complex mountainous terrain; the optimized deployment scheme includes: using an improved d-Xdraw visibility solution algorithm and three-dimensional visibility analysis to pre-calculate the effective coverage range of the sensor under fluctuating terrain, and combining a deployment solver based on an improved genetic algorithm or quantum particle swarm optimization algorithm to determine the optimal spatial layout position and attitude parameters of the sensor, thereby realizing the maximization of coverage of the monitoring area and the continuity and integrity of data acquisition.

[0047] S2: multi-dimensional data preprocessing and feature enhancement step:

[0048] This step performs a series of preprocessing operations on the collected multi-modal data, including:

[0049] Optical image radiometric calibration and atmospheric correction.

[0050] Geometric correction and orthorectification.

[0051] Multi-temporal image registration.

[0052] Image registration employs an automatic registration method based on SIFT features and the RANSAC algorithm, with registration errors controlled within one pixel. The registered images then undergo multi-level data fusion.

[0053] F pixel =α·I optical +(1-α)·I InSAR

[0054] Among them, F pixel For pixel-level fusion results, I optical and I InSAR These are optical and radar images, respectively, with α being the fusion weighting coefficient (0.6-0.8).

[0055] The multidimensional data preprocessing and feature enhancement steps also include multi-level data fusion processing; this data fusion processing is performed before the data is input into the hybrid deep learning model, and specifically includes:

[0056] S21: Pixel-level fusion: Fusion of registered multi-source images at the pixel level to generate a new image with richer texture and spectral information;

[0057] S22: Feature-level fusion: Extract independent feature vectors from data from different sensors, and then concatenate or weightedly combine these feature vectors to form a multi-dimensional feature set with discriminative power;

[0058] S23: Decision-level fusion: Using methods such as DS evidence theory or Bayesian networks, the preliminary classification results based on a single data source are comprehensively evaluated to resolve potential conflicts and uncertainties between different data sources and generate reliable fusion decision results.

[0059] S3: Steps for intelligent recognition of deformable fields driven by hybrid deep learning models:

[0060] The hybrid deep learning model consists of two parts: CNN and LSTM.

[0061] The CNN module uses the ResNet50 architecture as its backbone network, taking a 256×256 pixel image patch as input and outputting a 512-dimensional feature vector. The loss function for the CNN is:

[0062] L CNN =-∑ i＝1 N y i log(pi )+ λ||w||2 2

[0063] where y i is the true label, p i is the predicted probability, and λ is the regularization coefficient (default 0.001).

[0064] The LSTM module consists of 3 layers of LSTM units with 128 hidden units in each layer, and the input is the sequence of time-sequential features extracted by CNN. The state update formula of LSTM is:

[0065] f t = σ(W f · [h t-1 , x t ] + b f )

[0066] i t = σ(W i · [h t-1 , x t ] + b i )

[0067] o t = σ(W o · [h t-1 , x t ] + b o )

[0068] c t = tanh(W c · [h t-1 , x t ] + b c )

[0069] c t = f t ° c t-1 + i t ° c t

[0070] h t = o t ° tanh(c t )

[0071] Before training, the convolutional neural network (CNN) module of the hybrid deep learning model is pre-trained using a visual terrain classification method based on multiple encoding and feature fusion; the pre-training process uses a large-scale geological image dataset, extracts SIFT features and performs PCA-Whitening preprocessing, combines GIST global features, constructs a multi-feature dictionary, and performs feature fusion through multi-kernel learning (MKL) to train a linear support vector machine (SVM) classifier, and initializes the weights of the SVM or the bottom convolution kernel of the CNN.

[0072] The FPGA is designed and implemented with a deep-optimized hardware pipeline architecture to implement the convolutional neural network (CNN) calculation in step S3; the pipeline architecture decomposes complex convolution operations, pooling operations, and activation function calculations into multiple fine and continuous execution sub-stages, performs calculations on different regions of the input feature map simultaneously through task parallelism and data parallelism, and fuses the linear transformation parameters of the BatchNorm layer and the convolution layer at the hardware level.

[0073] An efficient memory access optimization strategy is integrated in the hardware pipeline architecture of the FPGA; the strategy includes:

[0074] S31: Data prefetching and weight caching: while the current convolution layer calculation is being performed, a dedicated hardware prefetching logic is used to load the convolution kernel weights and part of the input feature map data required for the next layer calculation from the off-chip DDR to the high-speed block random access memory (BRAM) on the FPGA chip for caching;

[0075] The CNN hardware interface circuit implemented on the FPGA has its module function definitions and working timing achieved through a global synchronous clock and a hierarchical synchronization mechanism for hardware synchronization; the synchronization mechanism includes:

[0076] S311: Interface layer synchronization: using the handshaking mechanism provided by the AXI bus protocol, the synchronization of data transmission between the DMA controller and the BRAM cache area is ensured;

[0077] S312: Inter-module synchronization: within the FPGA, a central control module is designed to generate global synchronization clock signals and enable signals, which are distributed to the processing elements (PEs), cache controllers, and data prefetching modules in the pipeline, ensuring that all modules trigger operations at the same clock edge;

[0078] S313: Cross-clock domain synchronization: for the possible different clock domains, an asynchronous FIFO is used as a data buffer and synchronization bridge to safely transfer data.

[0079] S32: Double buffering ping-pong operation: two independent BRAMs are configured as double buffer for input feature map and output feature map respectively.

[0080] S4: Dynamic threshold classification and determination step of deformation type and risk level

[0081] This step uses a four-parameter logistic regression model to dynamically adjust the classification threshold:

[0082] T(t) = A + (B-A) / (1+(t / C) D )

[0083] Where T(t) is the threshold value at time t, A is the lower limit value, B is the upper limit value, C is the inflection point time, and D is the curve slope.

[0084] The geological state conversion function δ(S, I)→S' is defined as:

[0085] δ(S, I) =

[0086] S stable ifV max ≤T1andD acc ≤T2

[0087] S creep ifT1<V max ≤T3orT2<D acc ≤T4

[0088] S accel ifT3<V max ≤T5orT4<D acc ≤T6

[0089] S failure ifV max >T5orD acc >T6

[0090] FPGA hardware acceleration implementation

[0091] The hardware pipeline architecture of FPGA to implement CNN calculation includes:

[0092] Convolution calculation unit (PE array)

[0093] Double buffering BRAM memory

[0094] AXI-DMA data transmission controller

[0095] The convolution operation adopts Winograd fast algorithm, and the calculation complexity is reduced from O(n 2 ) to O(n 1.58). The calculation formula of a single PE is:

[0096] y = f(∑ i＝1 k ∑ j＝1 k w ij x ij +b)

[0097] where f(·) is a ReLU activation function, w ij is a convolution kernel weight, x ij is input feature map data, and b is a bias term.

[0098] The memory access optimization adopts a double-buffering ping-pong operation, the data transmission bandwidth utilization rate is more than 90%, and the calculation efficiency is increased by 3-5 times.

[0099] The dynamic threshold self-adaptive adjustment mechanism is a mathematical framework based on state conversion logic, which is used to accurately distinguish between landslide and ground subsidence and other geological deformation types; the framework defines at least four geological states: stable state (S_stable), initial creep state (S_creep), accelerated deformation state (S_accel) and unstable state (S_failure); the conversion between states is driven by a conversion function delta (S, I) -> S', wherein the current state S and the input vector I jointly determine the next state S'; the input vector I includes quantitative parameters extracted from the deformation field map; the threshold value for state conversion is not a fixed value, but is adjusted in real time by a dynamic function based on a four-parameter logistic regression model or Bayesian update, which can dynamically update the threshold value according to the statistical variability of historical monitoring data, thereby improving the accuracy of classification and the timeliness of early warning.

[0100] The application provides a machine vision recognition geological deformation method, which is executed by a heterogeneous computing platform integrating a central processing unit (CPU) and a field programmable gate array (FPGA); wherein the control logic and complex decision involved in data acquisition (S1), data preprocessing (S2) and dynamic threshold classification (S4) are responsible by the CPU, and the most computationally intensive deformation field intelligent recognition step (S3) is completely offloaded to the FPGA for hardware acceleration.

[0101] In order to cope with the needs of CNN model structure change or multi-task parallel processing, a hardware task scheduling algorithm based on reconfigurable computing is realized on the FPGA; the scheduling algorithm comprehensively evaluates the task correlation, resource utilization and internal communication overhead according to the CNN network topology structure, intelligently maps and layouts the calculation tasks with close communication to the physically adjacent FPGA logic resources, and adjusts the hardware function at runtime through dynamic partial reconfiguration (DPR) technology.

[0102] The application provides a machine vision identification geological deformation method, multi-sensor optimal deployment solves the complex terrain signal shielding and data blind area problem, realizes monitoring area maximization coverage and data acquisition continuous integrity. Multidimensional data preprocessing and multilevel fusion processing effectively correct data distortion, eliminate noise, highlight weak features, improve feature extraction recognition degree and classification result reliability. The hybrid deep learning model couples CNN and LSTM, can extract spatial structure features and capture dynamic process at the same time, output high resolution, high precision geological deformation field map. Dynamic threshold self-adaptive adjustment mechanism accurately distinguishes deformation type, quantifies risk grade, improves classification accuracy and early warning timeliness. Heterogeneous computing platform and FPGA hardware acceleration are adopted, which significantly reduces processing delay and meets the needs of quasi-real-time monitoring. The deep optimization hardware pipeline architecture, memory access optimization strategy, reconfigurable computing scheduling algorithm and hardware synchronization mechanism of FPGA improve data throughput, computing efficiency and system stability, and enhance the adaptability of hardware to model structure change.

[0103] The above-mentioned is only the embodiment of the application, and does not limit the patent range of the application, and any equivalent structure or equivalent process transformation using the content of the specification and drawings of the application, or directly or indirectly applied in other related technical fields, are also included in the patent protection range of the application.

Claims

1. A method of machine vision identification of geological deformation, characterized in that, Comprising the following steps: S1: Multi-source heterogeneous data periodic acquisition step: In the geological area to be monitored, a multi-sensor monitoring network composed of multiple types of sensors is deployed, and a predetermined acquisition period is set. The multi-sensor monitoring network at least includes an optical imaging device for obtaining high-resolution surface texture and morphology information, an interferometric synthetic aperture radar (InSAR) sensor for penetrating vegetation and clouds to obtain all-weather surface micro-deformation data, and a laser radar (LiDAR) sensor for generating a high-precision three-dimensional digital elevation model (DEM); the step aims to periodically acquire multi-modal, multi-temporal remote sensing observation data sets covering the entire monitoring area; S2: Multi-dimensional data preprocessing and feature enhancement step: The multi-modal, multi-temporal remote sensing observation data set collected in step S1 is preprocessed to correct data distortion introduced by sensors and environmental factors, eliminate noise interference, and highlight weak features related to geological deformation. The preprocessing operation specifically includes: performing radiation calibration and atmospheric correction on optical images to eliminate the effects of light and atmosphere; performing high-precision geometric correction and orthographic correction on all sensor data to eliminate geometric deformation caused by terrain undulations; and strictly registering images of different periods and different sensors to ensure spatial alignment at the pixel level; S3: Deformation field intelligent identification step driven by a hybrid deep learning model: The time series data set preprocessed in step S2 is input into a pre-trained hybrid deep learning model for geological deformation identification. The hybrid deep learning model couples a convolutional neural network (CNN) and a long short-term memory network (LSTM). The convolutional neural network (CNN) module is responsible for extracting spatial structural features such as cracks, settlement edges, and landslide body boundaries from single or differential images, generating feature maps representing deformation distribution. The long short-term memory network (LSTM) module is responsible for analyzing the sequence of feature maps output by the CNN module in chronological order to capture the nonlinear dynamic process and long-term dependence of geological deformation over time; S4: Dynamic threshold classification and determination step of deformation type and risk level: Based on the time series geological deformation field map generated in step S3, a dynamic threshold self-adaptive adjustment mechanism integrating state transition logic is used to finely divide the identified deformation area into different types of geological deformation and quantitatively evaluate the potential risk level. The dynamic threshold adjustment mechanism can automatically adjust the discrimination threshold for distinguishing different deformation states according to the real-time updated deformation field data and the pre-set physical model, and combine with the geological structure background knowledge base.

2. The method of machine vision identification of geological deformations according to claim 1, characterized by, The deployment scheme of the multi-sensor monitoring network in step S1 is optimized to solve the problem of signal shielding and data blind area in complex mountainous terrain; The optimization deployment scheme comprises: pre-calculating the effective coverage range of the sensor under the undulating terrain by using an improved d-Xdraw visual field solving algorithm and three-dimensional visibility analysis, and combining a deployment solver based on an improved genetic algorithm or a quantum particle swarm optimization algorithm to determine the optimal spatial layout position and attitude parameter of the sensor, so that the maximization coverage of the monitoring area and the continuity and integrity of data acquisition are realized.

3. The method of claim 1, wherein, The multi-dimensional data preprocessing and feature enhancement step in the step S2 further comprises a multi-level data fusion processing; the data fusion processing is performed before the data is input into the hybrid deep learning model, and specifically comprises: S21: pixel-level fusion: fusing the registered multi-source images at the pixel level to generate new images with richer texture and spectral information; S22: feature-level fusion: independently extracting feature vectors from different sensor data, and then splicing or weighted combining the feature vectors to form a discriminative multi-dimensional feature set; S23: decision-level fusion: using D-S evidence theory or Bayesian network and the like to comprehensively analyze and judge the preliminary classification results based on single data source, solve the possible conflicts and uncertainties between different data sources, and generate reliable fusion decision results.

4. The method of claim 1 or 3, wherein, Before training, the convolutional neural network (CNN) module of the hybrid deep learning model in the step S3 adopts a visual terrain classification method based on multi-encoding and feature fusion for pre-training; the pre-training process uses a large-scale geological image data set, extracts SIFT features and performs PCA-Whitening preprocessing, combines GIST global features, constructs a multi-feature dictionary, and performs feature fusion through multi-kernel learning (MKL) to train a linear support vector machine (SVM) classifier, and initializes the weights of the SVM or the bottom convolution kernel of the CNN.

5. The method of claim 1, wherein, The dynamic threshold self-adaptive adjustment mechanism in the step S4 is a mathematical framework based on state transition logic, which is used to accurately distinguish between landslide and ground subsidence and other geological deformation types; the framework defines at least four geological states: stable state (S_stable), initial creep state (S_creep), accelerated deformation state (S_accel) and unstable state (S_failure); the transition between states is driven by a transition function δ (S, I) → S', wherein the current state S and the input vector I jointly determine the next state S'; the input vector I includes quantitative parameters extracted from the deformation field map; the threshold value for state transition is not a fixed value, but is adjusted in real time by a dynamic function based on a four-parameter logistic regression model or Bayesian update, which can dynamically update the threshold value according to the statistical variability of historical monitoring data, thereby improving the accuracy of classification and the timeliness of early warning.

6. The method of machine vision identification of geological deformations according to claim 1, characterized by, The method is executed by a heterogeneous computing platform integrating a central processing unit (CPU) and a field programmable gate array (FPGA); wherein the control logic and complex decision involved in data acquisition (S1), data preprocessing (S2) and dynamic threshold classification (S4) are responsible by the CPU, and the most computationally intensive step of intelligent deformation field recognition (S3) is completely offloaded to the FPGA for hardware acceleration.

7. The method of machine vision identification of geological deformations according to claim 6, characterized by, The FPGA is designed and implemented with a deeply optimized hardware pipeline architecture for the convolutional neural network (CNN) calculation in step S3; The pipeline architecture decomposes complex convolution operation, pooling operation and activation function calculation into multiple fine and continuous executable sub-stages, performs calculation on different regions of the input feature map by multiple processing elements (PEs) through task parallelism and data parallelism, and fuses the linear transformation parameters of the BatchNorm layer and the convolution layer at the hardware level.

8. The method of machine vision identification of geological deformations according to claim 7, characterized by the fact that, An efficient memory access optimization strategy is integrated in the hardware pipeline architecture of the FPGA; the strategy includes: S31: data prefetching and weight caching: while the current convolution layer calculation is being executed, the convolution kernel weights and part of the input feature map data required for the next layer calculation are loaded from the off-chip DDR to the high-speed block random access memory (BRAM) on the FPGA chip for caching through a special hardware prefetching logic; S32: double-buffering ping-pong operation: two independent BRAMs are configured as double-buffering zones for input feature map and output feature map respectively; when the calculation core is reading data from one buffer for operation, the data transfer module.

9. The method of machine vision identification of geological deformations according to claim 7, characterized by the fact that, To cope with the requirement of CNN model structure change or multi-task parallel processing, a hardware task scheduling algorithm based on reconfigurable computing is implemented on the FPGA; the scheduling algorithm comprehensively evaluates the task correlation, resource utilization and internal communication overhead according to the CNN network topology, intelligently maps and lays out the calculation tasks with close communication to the physically adjacent FPGA logic resources, and adjusts the hardware function at runtime through dynamic partial reconfiguration (DPR) technology.

10. The method of machine vision identification of geological deformations according to claim 8, characterized by, The CNN hardware interface circuit implemented on the FPGA realizes hardware synchronization through a global synchronous clock and a hierarchical synchronization mechanism; the synchronization mechanism includes: S311: interface layer synchronization: the AXI bus protocol built-in handshaking mechanism is used to ensure the synchronization of data transmission between the DMA controller and the BRAM buffer; S312: inter-module synchronization: a central control module is designed inside the FPGA to generate global synchronization clock signals and enable signals, which are distributed to the processing elements (PEs), cache controllers and data prefetching modules in the pipeline, to ensure that all modules trigger operations at the same clock edge; S313: cross-clock domain synchronization: asynchronous FIFO is used as a data buffer and synchronization bridge for possible different clock domains to safely transfer data.

Images