A slope geological disaster AI early warning method and system based on an optical fiber sensing network

Abstract

This invention discloses an AI-based early warning method and system for slope geological disasters based on an optical fiber sensor network, belonging to the field of geological disaster monitoring and early warning technology. It constructs a full-coverage sensing network by embedding DAS optical fibers and laying DTS optical fibers on the surface, simultaneously collecting multi-source data on vibration and temperature. After preprocessing, the data is input into a convolutional neural network to automatically extract abnormal pattern features and accurately identify four types of signals: excavation, blasting, seepage, and micro-fractures in the rock mass. An entropy weighting method is used to weight and fuse multi-parameter features to establish a mapping relationship between disaster type and level, achieving a leap from simple alarm to accurate identification of disaster type and level. This invention solves the problems of insufficient sensing coverage, low identification accuracy, and vague warnings in traditional early warning methods, and has the advantages of comprehensive monitoring, accurate identification, and scientific early warning, making it suitable for geological disaster early warning on high-risk slopes.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring and early warning technology, and more specifically to an AI-based early warning method and system for slope geological disasters based on fiber optic sensor networks. Background Technology

[0002] Currently, slope geological disasters (such as landslides and collapses) are characterized by their suddenness and destructive power, seriously threatening engineering construction and the safety of people's lives and property. Traditional slope monitoring methods mainly rely on single-point sensor deployment, which has problems such as limited sensing range, unstable data transmission, and weak anti-interference ability. Early warning methods are mostly based on single parameter threshold judgments, which can only achieve simple alarms and cannot accurately identify the type and level of disasters, resulting in high false alarm and missed alarm rates.

[0003] Distributed fiber optic sensing technology offers advantages such as long sensing distance, distributed measurement, resistance to electromagnetic interference, and tolerance to harsh environments, enabling comprehensive slope monitoring. Artificial intelligence technology possesses powerful feature mining capabilities in pattern recognition, providing the possibility for accurate identification of abnormal patterns. However, current technologies have not effectively combined DAS / DTS full-dimensional perception with CNN pattern recognition and multi-parameter fusion early warning, lacking a unified mathematical model and standardized implementation process, thus limiting the accuracy and practicality of early warning systems.

[0004] Therefore, how to overcome the shortcomings of existing slope early warning methods, such as incomplete perception coverage, low feature recognition accuracy, unscientific multi-parameter fusion, and ambiguous early warning levels, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides an AI-based early warning method and system for slope geological disasters based on fiber optic sensor networks, in order to solve the problems existing in the background technology.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for AI-based early warning of slope geological hazards based on fiber optic sensor networks includes: Distributed acoustic sensing optical fibers are pre-buried and laid along the boreholes inside the high-risk slope, and distributed temperature sensing optical fibers are laid in a grid pattern along the slope surface to form a distributed sensing network for the slope. Vibration and temperature data from multiple sources are acquired through a distributed slope sensing network. The collected vibration and temperature multi-source data were preprocessed to obtain a standardized dataset; A convolutional neural network was used to extract features and recognize patterns from a standardized dataset to obtain abnormal pattern recognition results. Based on the results of abnormal pattern recognition, a multi-parameter weighted fusion early warning model is constructed, which integrates vibration characteristics and temperature characteristics, establishes a mapping relationship between disaster type and level, and outputs early warning results.

[0007] Optionally, the temperature data acquisition of the distributed temperature sensing fiber is based on the Raman scattering principle, and the measurement formula is:

[0008] Where h is Planck's constant and c is the speed of light in a vacuum. The frequency shift is represented by the Raman frequency shift, and k is the Boltzmann constant. Let z be the anti-Stokes light intensity at position z and time t. Let z be the Stokes light intensity at position z and time t. For anti-Stokes wavelength, The wavelength of Stokes light. Let z be the slope temperature at position z and time t.

[0009] Optionally, the vibration signal acquisition of the distributed acoustic sensing fiber is based on the Rayleigh scattering phase modulation principle, and the signal model is:

[0010] in, Let be the intensity of the scattered light at position z and time t. The average light intensity. The modulation coefficient, Let n be the initial phase and n be the refractive index of the fiber. Let z be the vibration displacement at position z and time t. The incident light wavelength, The phase represents random noise. Optionally, the data preprocessing employs Min-Max normalization, as shown in the formula:

[0011] in, For multi-source vibration and temperature data, For normalized data, The minimum value in the dataset. The maximum value in the dataset. Optionally, the convolutional neural network includes an input layer, two convolutional layers, two pooling layers, one fully connected layer, and an output layer; The input layer is the raw data receiver for the entire convolutional neural network, and the input data is a preprocessed standardized spatiotemporal dataset. The first convolutional layer is adjacent to the input layer. It performs preliminary feature extraction on the three-dimensional tensor data of the input layer and captures the basic low-dimensional features in the data. During the operation, each convolutional kernel performs two-dimensional convolution operation with the three-dimensional tensor of the input layer. At the same time, a bias term with an initial value of 0 is superimposed. Then, the convolution result is non-linearly transformed by the ReLU activation function to filter invalid features and enhance the discriminability of valid features. Finally, the feature map is output. The first pooling layer is adjacent to the first convolutional layer and performs dimensionality reduction and compression on the feature map output by the first convolutional layer. During the operation, the pooling window traverses each feature map output by the first convolutional layer with a fixed stride and takes the maximum value in each window as the feature value at the corresponding output position of that window, and finally compresses the feature map of the first convolutional layer into a dimensionality reduction feature map. The second convolutional layer is adjacent to the first pooling layer. It performs deep feature extraction on the dimensionality reduction feature map output by the first pooling layer to capture high-order complex features in the data. The second pooling layer is adjacent to the second convolutional layer and has the same function as the first pooling layer, which is dimensionality reduction and compression. The fully connected layer is adjacent to the second pooling layer, which converts the high-dimensional feature map output by the second pooling layer into a one-dimensional feature vector with fixed dimensions; The output layer is the output of the entire convolutional neural network. It is adjacent to the fully connected layer and performs pattern classification on the feature vectors output by the fully connected layer to identify the state category corresponding to the data.

[0012] Optionally, the input layer dimension of the convolutional neural network is 64×64×2, the number of filters in the first convolutional layer is 32, the number of filters in the second convolutional layer is 64, the output dimension of the fully connected layer is 128, and the output layer uses the Softmax activation function.

[0013] Optionally, the multi-parameter weighted fusion early warning model uses the entropy weight method to determine the weights, and the fusion formula is as follows: Feature weight calculation:

[0014] in, For the first i The weights of class features, i=v for vibration features, i=t for temperature features; For information entropy, For the first i The normalized proportion of the m-th sample of a class feature. These are the original eigenvalues; Fusion feature calculation:

[0015] Where F is the fusion feature value, The vibration feature vector is the output of the convolutional neural network. This represents the temperature feature vector. A slope geological disaster AI early warning system based on fiber optic sensor networks includes: The sensing network setup module involves pre-burying distributed acoustic sensing optical fibers along the boreholes inside the high-risk slope and laying distributed temperature sensing optical fibers in a grid pattern along the slope surface to form a distributed sensing network for the slope. The data acquisition module acquires multi-source vibration and temperature data through a distributed slope sensing network. The data preprocessing module preprocesses the collected vibration and temperature multi-source data to obtain a standardized dataset; The abnormal result identification module uses a convolutional neural network to extract features and recognize patterns from a standardized dataset to obtain abnormal pattern identification results. The early warning result output module constructs a multi-parameter weighted fusion early warning model based on the abnormal pattern recognition results, integrates vibration characteristics and temperature characteristics, establishes a mapping relationship between disaster type and level, and outputs early warning results.

[0016] As can be seen from the above technical solution, compared with the prior art, the present invention discloses an AI early warning method and system for slope geological disasters based on fiber optic sensor networks, which has the following beneficial effects: 1. This invention pre-embeds distributed acoustic sensing optical fibers along the boreholes inside high-risk slopes and lays distributed temperature sensing optical fibers in a grid pattern along the slope surface to form a distributed slope sensing network, thereby realizing distributed monitoring of the slope in three-dimensional space, with no monitoring blind spots and improving the comprehensiveness of sensing coverage. 2. This invention is based on a deep learning algorithm of convolutional neural networks, which automatically mines multi-dimensional abnormal features, improves the recognition accuracy, and effectively distinguishes between interference signals and disaster signals; 3. This invention uses the entropy weight method to determine feature weights, which conforms to the physical mechanism of disaster evolution and makes the fused features more representative; 4. This invention enables dual early warning of disaster type and risk level, breaking through the limitations of traditional simple alarms and providing accurate basis for emergency response. Attached Figure Description

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

[0018] Figure 1 This is a schematic diagram of the method flow provided by the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention discloses an AI-based early warning method for slope geological disasters based on fiber optic sensor networks, such as... Figure 1 As shown, it includes: Distributed acoustic sensing (DAS) optical fibers are pre-embedded and laid along the boreholes inside the high-risk slope, and distributed temperature sensing (DTS) optical fibers are laid in a grid pattern along the slope surface to form a distributed sensing network for the slope. Vibration and temperature data from multiple sources are acquired through a distributed slope sensing network. The collected vibration and temperature multi-source data were preprocessed to obtain a standardized dataset; A convolutional neural network was used to extract features and recognize patterns from a standardized dataset to obtain abnormal pattern recognition results. Based on the results of abnormal pattern recognition, a multi-parameter weighted fusion early warning model is constructed, which integrates vibration characteristics and temperature characteristics, establishes a mapping relationship between disaster type and level, and outputs early warning results.

[0021] Specifically, the distributed sensing network for slopes is constructed as follows: DAS fiber optic deployment: Drill holes in the potential sliding surface area of ​​the slope, encapsulate the DAS fiber optic cable in a stainless steel sleeve and pre-embed it. The deployment depth is 0.5-1.0m below the sliding surface, and the lateral spacing is ≤5m to ensure coverage of key monitoring areas. DTS fiber optic deployment: The fiber optic cable is laid in a grid pattern on the slope surface with a grid density of 2m×2m. The fiber optic cable is fixed to the slope surface with fixing clips and a bending radius of ≥30cm to avoid signal attenuation. Data acquisition system: uses pulsed laser light source, DAS sampling rate 1kHz, spatial resolution 1m; DTS sampling interval 0.5m, temperature resolution 0.1℃.

[0022] In one specific embodiment, the temperature data acquisition of the distributed temperature sensing fiber is based on the Raman scattering principle, and the measurement formula is:

[0023] Where h is Planck's constant and c is the speed of light in a vacuum. The frequency shift is represented by the Raman frequency shift, and k is the Boltzmann constant. Let z be the anti-Stokes light intensity at position z and time t. Let z be the Stokes light intensity at position z and time t. For anti-Stokes wavelength, The wavelength of Stokes light. Let z be the slope temperature at position z and time t.

[0024] In one specific embodiment, the vibration signal acquisition of the distributed acoustic sensing fiber is based on the Rayleigh scattering phase modulation principle, and the signal model is as follows:

[0025] in: Let be the intensity of the scattered light at position z and time t. The average light intensity. The modulation coefficient, Let n be the initial phase and n be the refractive index of the fiber. Let z be the vibration displacement at position z and time t. The incident light wavelength, This represents the phase of random noise. In a specific embodiment, data preprocessing employs Min-Max normalization to map temperature and vibration data to the [0,1] interval, eliminating dimensional differences. The formula is:

[0026] in, For multi-source vibration and temperature data, For normalized data, The minimum value in the dataset. The maximum value in the dataset. In one specific embodiment, the convolutional neural network includes an input layer, two convolutional layers, two pooling layers, one fully connected layer, and an output layer; The input layer is the raw data receiver for the entire convolutional neural network. The input data is a preprocessed standardized spatiotemporal dataset, with the network input dimension set to 64×64×2. Here, 64 represents the time step and the number of spatial monitoring points, respectively, and 2 represents two types of input features: vibration data acquired by distributed acoustic sensors and temperature data acquired by distributed temperature sensors. This layer reconstructs the standardized data into a three-dimensional tensor form, providing regular input data for subsequent feature extraction. It is the basic data source for convolution operations and corresponds to the original input parameters in the formulas of subsequent layers.

[0027] The first convolutional layer is adjacent to the input layer. Its core function is to perform preliminary feature extraction on the 3D tensor data of the input layer, capturing basic low-dimensional features in the data. This layer uses 3×3 convolutional kernels as the core tool for feature extraction, with 32 kernels in total. Each kernel generates one feature map. During the operation, each kernel performs a 2D convolution operation with the 3D tensor of the input layer, while simultaneously adding a bias term with an initial value of 0. Then, the ReLU activation function is used to perform a non-linear transformation on the convolution result, filtering out invalid features and enhancing the discriminative power of valid features, ultimately outputting a 32-dimensional 64×64 feature map. Its operational logic perfectly matches the general formula for convolutional layers:

[0028] In the formula, =1, the input layer contains only one set of three-dimensional tensor data. The input layer consists of three-dimensional tensor data. The first convolutional layer has 32 3×3 convolutional kernels. These are the 32 feature maps output by the first convolutional layer.

[0029] The first pooling layer, adjacent to the first convolutional layer, functions primarily to reduce the dimensionality of the feature map output by the first convolutional layer. This reduces the number of network parameters, improves computational efficiency, and avoids overfitting while preserving core, effective features. This layer employs max pooling, setting the pooling window size to s=2 (i.e., a 2×2 pooling window). During computation, the 2×2 pooling window iterates through each feature map output by the first convolutional layer with a fixed stride, taking the maximum value within each window as the feature value at the corresponding output position. Ultimately, this compresses the 64×64-dimensional feature map of the first convolutional layer into a 32×32-dimensional map, outputting 32 32×32-dimensional reduced feature maps. Its computational logic perfectly matches the max pooling layer formula.

[0030] In the formula, s=2 is the pooling window size. Let (p, q) be a single feature map output by the first convolutional layer, where (p, q) are the coordinates of the feature map output by this layer. To output the feature values ​​corresponding to the coordinates of the feature map, a is the horizontal local coordinate index within the pooling window, corresponding to the width dimension of the feature map, and b is the vertical local coordinate index within the pooling window, corresponding to the height dimension of the feature map.

[0031] The second convolutional layer is adjacent to the first pooling layer. Its core function is to extract deep features from the dimensionality-reduced feature map output by the first pooling layer, capturing high-order complex features in the data. This layer uses 3×3 convolutional kernels, but the number of kernels is increased to 64. Each kernel generates one high-order feature map. The operation logic is the same as the first convolutional layer. Each kernel performs a two-dimensional convolution operation with the 32 dimensionality-reduced feature maps output by the first pooling layer, and a bias term with an initial value of 0 is added. Then, after a non-linear transformation using the ReLU activation function, the final output is a high-order feature map with 64 dimensions of 32×32. Its computational logic also matches the general formula for convolutional layers:

[0032] In the formula, =32 (The first pooling layer outputs 32 feature maps). This is a single dimensionality-reduced feature map output by the first pooling layer. The second convolutional layer has 64 3×3 convolutional kernels. This is a single high-order feature map output by the second convolutional layer.

[0033] The second pooling layer is adjacent to the second convolutional layer and has the same function as the first pooling layer: dimensionality reduction and compression, preservation of core features, and optimization of network performance. This layer also uses max pooling, with the pooling window size remaining at s=2. During operation, the 2×2 pooling window traverses the 64 32×32-dimensional feature maps output by the second convolutional layer, taking the maximum value within each window as the output feature value, thus compressing the feature map dimension from 32×32 to 16×16, and finally outputting 64 16×16-dimensional depth-reduced feature maps.

[0034] The fully connected layer, immediately adjacent to the second pooling layer, functions primarily to convert the high-dimensional feature map output by the second pooling layer into a fixed-dimensional one-dimensional feature vector, thus facilitating the transition from feature extraction to pattern classification. This layer first flattens the 64 16×16 dimensional feature maps output by the second pooling layer, converting them into a single one-dimensional vector, and then applies a 128×256 weight matrix. Perform linear operations with the flattened vector, and then add a fully connected layer bias term with an initial value of 0. The data is then nonlinearly optimized using the ReLU activation function, ultimately outputting a one-dimensional feature vector with dimensions 128. This vector serves as the core high-dimensional representation of the data and is used for subsequent pattern recognition. Its computational logic perfectly matches the formula for a fully connected layer.

[0035] In the formula, This is a one-dimensional vector flattened from the second pooling layer. This is the 128-dimensional feature vector output by this layer.

[0036] The output layer, adjacent to the fully connected layer, is the output of the entire convolutional neural network. Its core function is to perform pattern classification on the 128-dimensional feature vector output by the fully connected layer, accurately identifying the state category corresponding to the data. This layer has 5 neurons, corresponding to 5 target categories: normal state, excavation anomaly, blasting anomaly, seepage anomaly, and rock micro-fracture anomaly. The Softmax activation function is used to convert the feature vector output by the fully connected layer into predicted probabilities for the 5 categories. The category with the highest probability is output as the pattern recognition result of the network. This recognition result, together with the 128-dimensional feature vector output by the fully connected layer, is fed into the subsequent multi-parameter fusion early warning model. The number of categories corresponds to the number of categories C=5 in the loss function, specifically including 4 abnormal modes: excavation, blasting, seepage, rock micro-fracture, and normal mode. To ensure the pattern recognition accuracy of convolutional neural networks, a loss function is used to measure the error between the predicted results and the true labels. Then, an optimizer iteratively updates the network parameters until the error is minimized and the model converges.

[0037] The loss function quantifies the deviation between the predicted probability output by the network and the true label of the sample. The label uses one-hot encoding, and the formula is:

[0038] In the formula, N is the total number of training samples, and C=5 is the number of target categories. Let m be the true label of the m-th sample. Let L be the predicted probability of the m-th sample corresponding to the c-th category, and L be the average loss value of a single training session. The smaller L is, the smaller the prediction bias.

[0039] The optimizer iteratively updates all learnable parameters of the network based on the gradient of the loss function, thereby minimizing the loss function L.

[0040] In a specific embodiment, the multi-parameter weighted fusion early warning model uses the entropy weight method to determine the weights, and the fusion formula is as follows: Specifically, the input data is used to obtain two types of core data from the output of the convolutional neural network: ① the abnormal pattern classification result C (C=1 for excavation, C=2 for blasting, C=3 for seepage, and C=4 for rock micro-fractures); ② the vibration feature vector of the corresponding pattern.

[0041] Synchronous data is obtained from the distributed temperature sensing (DTS) system: the raw temperature data after Min-Max normalization is reconstructed through the same dimension as the vibration feature to obtain the temperature feature vector, ensuring that the temperature feature vector and the vibration feature vector have the same dimension and can be directly fused.

[0042] For a specific anomalous pattern C identified by the CNN, the historical feature dataset corresponding to that pattern is accessed, and the specific weights of vibration and temperature features under that pattern are calculated using the entropy weight method.

[0043] in, For the first i The weights of class features, i=v for vibration features, i=t for temperature features; For information entropy, For the first i The normalized proportion of the m-th sample of a class feature. The original feature values ​​are represented by N; N is the total number of samples; k is the index variable, which is calculated by iterating through v and t using k. This yields the weighted denominator; Fusion feature calculation:

[0044] in, The vibration feature vector is the output of the convolutional neural network. is the temperature feature vector. F is the fused feature value, which comprehensively reflects the risk intensity under the current abnormal mode.

[0045] For the abnormal pattern C identified by the convolutional neural network, the preset warning threshold corresponding to the pattern is called (T1 is the high-risk threshold, T2 is the medium-risk threshold). The fused feature value F is compared with the threshold, and the risk level is determined according to the following rules to establish a precise mapping from disaster type C to risk level:

[0046] By linking and integrating the results of abnormal pattern recognition with the results of risk level determination, a complete early warning information is formed, which includes disaster type: clearly identified abnormal patterns by CNN, and risk level: clearly determined risk level.

[0047] The early warning system interface displays the disaster type, risk level, location of occurrence, and characteristic value F. Tiered alarms: Level 1 risk triggers an audible and visual alarm and sends an SMS to management personnel; Level 2 risk triggers an audible and visual alarm; Level 3 risk only provides an on-screen notification.

[0048] An AI-based early warning system for slope geological hazards based on fiber optic sensor networks includes: The sensing network setup module involves pre-burying distributed acoustic sensing optical fibers along the boreholes inside the high-risk slope and laying distributed temperature sensing optical fibers in a grid pattern along the slope surface to form a distributed sensing network for the slope. The data acquisition module acquires multi-source vibration and temperature data through a distributed slope sensing network. The data preprocessing module preprocesses the collected vibration and temperature multi-source data to obtain a standardized dataset; The abnormal result identification module uses a convolutional neural network to extract features and recognize patterns from a standardized dataset to obtain abnormal pattern identification results. The early warning result output module constructs a multi-parameter weighted fusion early warning model based on the abnormal pattern recognition results, integrates vibration characteristics and temperature characteristics, establishes a mapping relationship between disaster type and level, and outputs early warning results.

[0049] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0050] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for AI-based early warning of slope geological hazards based on fiber optic sensor networks, characterized in that, include: Distributed acoustic sensing optical fibers are pre-buried and laid along the boreholes inside the high-risk slope, and distributed temperature sensing optical fibers are laid in a grid pattern along the slope surface to form a distributed sensing network for the slope. Vibration and temperature data from multiple sources are acquired through a distributed slope sensing network. The collected vibration and temperature multi-source data were preprocessed to obtain a standardized dataset; A convolutional neural network was used to extract features and recognize patterns from a standardized dataset to obtain abnormal pattern recognition results. Based on the results of abnormal pattern recognition, a multi-parameter weighted fusion early warning model is constructed, which integrates vibration characteristics and temperature characteristics, establishes a mapping relationship between disaster type and level, and outputs early warning results.

2. The AI-based early warning method for slope geological disasters based on fiber optic sensor networks according to claim 1, characterized in that, The temperature data acquisition of the distributed temperature sensing fiber is based on the Raman scattering principle, and the measurement formula is: Where h is Planck's constant and c is the speed of light in a vacuum. The frequency shift is represented by the Raman frequency shift, and k is the Boltzmann constant. Let z be the anti-Stokes light intensity at position z and time t. Let z be the Stokes light intensity at position z and time t. For anti-Stokes wavelength, The wavelength of Stokes light. Let z be the slope temperature at position z and time t.

3. The AI-based early warning method for slope geological disasters based on fiber optic sensor networks according to claim 1, characterized in that, The vibration signal acquisition of the distributed acoustic sensing fiber is based on the Rayleigh scattering phase modulation principle, and the signal model is as follows: in, Let be the intensity of the scattered light at position z and time t. The average light intensity The modulation coefficient, Let n be the initial phase and n be the refractive index of the fiber. Let z be the vibration displacement at position z and time t. The incident light wavelength, It represents random noise phase.

4. The AI-based early warning method for slope geological disasters based on fiber optic sensor networks according to claim 1, characterized in that, The data preprocessing uses Min-Max normalization, with the following formula: in, For multi-source vibration and temperature data, For normalized data, The minimum value in the dataset. This represents the maximum value in the dataset.

5. The AI-based early warning method for slope geological disasters based on fiber optic sensor networks according to claim 1, characterized in that, The convolutional neural network includes an input layer, two convolutional layers, two pooling layers, one fully connected layer, and an output layer. The input layer is the raw data receiver for the entire convolutional neural network, and the input data is a preprocessed standardized spatiotemporal dataset. The first convolutional layer is adjacent to the input layer. It performs preliminary feature extraction on the three-dimensional tensor data of the input layer and captures the basic low-dimensional features in the data. During the operation, each convolutional kernel performs two-dimensional convolution operation with the three-dimensional tensor of the input layer. At the same time, a bias term with an initial value of 0 is superimposed. Then, the convolution result is non-linearly transformed by the ReLU activation function to filter invalid features and enhance the discriminability of valid features. Finally, the feature map is output. The first pooling layer is adjacent to the first convolutional layer and performs dimensionality reduction and compression on the feature map output by the first convolutional layer. During the operation, the pooling window traverses each feature map output by the first convolutional layer with a fixed stride and takes the maximum value in each window as the feature value at the corresponding output position of that window, and finally compresses the feature map of the first convolutional layer into a dimensionality reduction feature map. The second convolutional layer is adjacent to the first pooling layer. It performs deep feature extraction on the dimensionality reduction feature map output by the first pooling layer to capture high-order complex features in the data. The second pooling layer is adjacent to the second convolutional layer and has the same function as the first pooling layer, which is dimensionality reduction and compression. The fully connected layer is adjacent to the second pooling layer, which converts the high-dimensional feature map output by the second pooling layer into a one-dimensional feature vector with fixed dimensions; The output layer is the output of the entire convolutional neural network. It is adjacent to the fully connected layer and performs pattern classification on the feature vectors output by the fully connected layer to identify the state category corresponding to the data.

6. The AI-based early warning method for slope geological disasters based on fiber optic sensor networks according to claim 1, characterized in that, The input layer of the convolutional neural network has a dimension of 64×64×2, the first convolutional layer has 32 filters, the second convolutional layer has 64 filters, the fully connected layer has an output dimension of 128, and the output layer uses the Softmax activation function.

7. The AI-based early warning method for slope geological disasters based on fiber optic sensor networks according to claim 1, characterized in that, The multi-parameter weighted fusion early warning model uses the entropy weight method to determine the weights, and the fusion formula is as follows: Feature weight calculation: in, For the first i The weights of class features, i=v for vibration features, i=t for temperature features; For information entropy, For the first i The normalized proportion of the m-th sample of a class feature. These are the original eigenvalues; Fusion feature calculation: Where F is the fusion feature value, The vibration feature vector is the output of the convolutional neural network. This is the temperature feature vector.

8. A slope geological disaster AI early warning system based on fiber optic sensor network, characterized in that, The method for AI-based early warning of slope geological hazards based on fiber optic sensor networks, as described in any one of claims 1-7, includes: The sensing network setup module involves pre-burying distributed acoustic sensing optical fibers along the boreholes inside the high-risk slope and laying distributed temperature sensing optical fibers in a grid pattern along the slope surface to form a distributed sensing network for the slope. The data acquisition module obtains multi-source vibration and temperature data through a distributed slope sensing network; The data preprocessing module preprocesses the collected vibration and temperature multi-source data to obtain a standardized dataset; The abnormal result identification module uses a convolutional neural network to extract features and recognize patterns from a standardized dataset to obtain abnormal pattern identification results. The early warning result output module constructs a multi-parameter weighted fusion early warning model based on the abnormal pattern recognition results, integrates vibration characteristics and temperature characteristics, establishes a mapping relationship between disaster type and level, and outputs early warning results.

Images