Geological disaster early warning algorithm based on hyperspectrum and Internet of Things data fusion analysis

Abstract

The invention relates to a geological disaster early warning algorithm based on hyperspectral and Internet of Things data fusion analysis, and relates to the technical field of geological disaster monitoring and early warning, the geological disaster early warning algorithm comprises the following steps: S1, preprocessing hyperspectral data and Internet of Things data, and respectively extracting spectral-spatial features and dynamic topology time sequence features; s2, fusing the spectrum-space features and the time sequence features to generate cross-modal joint features; s3, performing parameter optimization on the fusion features through a quantum-classical hybrid optimization algorithm, and calculating a disaster risk probability; and S4, based on the optimization result and the risk probability, executing an edge-cloud collaborative early warning decision. According to the method, through non-negative tensor ring decomposition of the hyperspectral data and dynamic topology modeling of the Internet of Things, the limitation of a traditional single data source in temporal-spatial resolution and physical relevance is solved, multi-dimensional joint extraction of spectrum-space-mechanical characteristics can be realized, and the characterization precision of a rock-soil body deformation evolution law is remarkably enhanced.

Description

Technical Field

[0001] This application relates to the field of geological disaster monitoring and early warning technology, and in particular to a geological disaster early warning algorithm based on the fusion analysis of hyperspectral and Internet of Things data. Background Technology

[0002] Current geological disaster early warning systems mainly rely on a single data source to build analytical models, and this approach has significant limitations in complex geological environments. Although hyperspectral remote sensing technology can acquire spectral-spatial multidimensional information of surface materials, traditional dimensionality reduction methods (such as principal component analysis and independent component analysis) are prone to destroying the nonlinear correlation between spectral bands when processing high-dimensional data, resulting in insufficient accuracy in lithology identification and deformation feature extraction.

[0003] Meanwhile, ground-based IoT sensor networks are affected by factors such as dynamic changes in ground stress, node communication interference, and fluctuations in data confidence. Static graph neural networks struggle to effectively model the dynamic mechanical relationships between sensor nodes, leading to delays or misjudgments in the detection of time-series anomalies. Existing data-driven models often employ pure black-box deep learning architectures, which, while possessing high fitting capabilities in specific scenarios, lack explicit embedding of fundamental laws of geotechnical mechanics (such as stress balance equations and constitutive relations). Consequently, prediction results may deviate from physical reality, reducing their engineering guidance value.

[0004] Furthermore, traditional parameter optimization methods (such as gradient descent and genetic algorithms) are easily limited by local optima and convergence speed when dealing with high-dimensional tensor decomposition and multi-objective constrained problems, making it difficult to meet the computational efficiency requirements for real-time early warning under complex geological conditions. At the system architecture level, existing solutions often separate real-time inference at the edge from model training in the cloud, making it difficult to balance real-time response speed and long-term prediction accuracy in resource-constrained scenarios. Moreover, there is a risk of privacy leakage during multi-node data collaborative training, which restricts the actual deployment effectiveness of large-scale geological disaster monitoring networks. Summary of the Invention

[0005] The purpose of this application is to provide a geological disaster early warning algorithm based on the fusion analysis of hyperspectral and Internet of Things data, which solves the problems of insufficient handling of multimodal data heterogeneity, limited model generalization ability, and lack of physical interpretability of early warning results in traditional geological disaster early warning methods.

[0006] Firstly, the geological disaster early warning algorithm based on the fusion analysis of hyperspectral and Internet of Things data provided in this application adopts the following technical solution: Geological disaster early warning algorithms based on the fusion analysis of hyperspectral and IoT data include: S1. Preprocess hyperspectral data and IoT data to extract spectral-spatial features and dynamic topological temporal features, respectively; S2. Fuse the spectral-spatial features and temporal features to generate cross-modal joint features; S3. Optimize the parameters of the fused features using a quantum-classical hybrid optimization algorithm, and calculate the probability of disaster risk; S4. Based on the optimization results and risk probability, execute edge-cloud collaborative early warning decisions.

[0007] Preferably, the preprocessing of hyperspectral data in step S1 includes: in, To input a hyperspectral cube, Let R1, ..., R be the core tensor after decomposition. N Let be the rank parameters; nonlinear correlations between rank parameters are established through quantum entanglement encoding, specifically mapped as follows: Where q1 and q2 are qubits, and |ψ> represents an entangled state.

[0008] Preferably, the preprocessing of IoT data in step S1 includes: The dynamic adjacency matrix of sensor nodes is constructed, and its weight calculation expression is as follows: Where F i Let i be the shear vector at node i. i Let σ and α be the data confidence level, and α be learnable parameters; temporal features are extracted using a dynamic graph neural network, and its update formula is: H (l+1) =ReLU(AH (l) W (l) ); Where H (l) For the features of the l-th layer nodes, W (l) This is a trainable weight matrix.

[0009] Preferably, the cross-modal joint feature fusion in step S2 includes: Hyperspectral tensor features and IoT temporal features are concatenated along the parameter dimension to form a multidimensional tensor; Separable tensor convolution kernels are used for cross-modal feature interaction computation.

[0010] Preferably, the fusion process further includes: The constraint of the geotechnical equilibrium equation is embedded in the loss function, and its expression is as follows: Where σ t To predict the stress tensor, b tIt is a volume force vector. It is a divergence operator.

[0011] Preferably, the quantum-classical hybrid optimization in step S3 includes: Mapping the tensor decomposition problem to the quantum Ising model, its Hamiltonian expression is: H = -∑ i<j J ij q i q j -∑ i h i q i ; Where, q i ∈{0,1} represents the rank parameter selection state, J ij with h i Calculated from tensor reconstruction error; The quantum annealing results are used as initialization parameters for a classical neural network, and jointly optimized through a gradient sharing mechanism.

[0012] Preferably, the quantum optimization further includes: Encoding data uncertainty as quantum phase noise, its quantum state expression is as follows: Where ∈ represents data uncertainty, U quantum For quantum circuit unitary operators.

[0013] Preferably, the edge-cloud collaborative early warning decision includes: A binary quantization model is deployed at the edge for real-time anomaly detection, and its weight quantization formula is as follows: Where n is the total number of weights, and Δ is the dynamic scaling factor; The geotechnical constitutive equations are dynamically updated in the cloud based on real-time data.

[0014] Preferably, the cloud update further includes: By aggregating parameters from multiple edge nodes using a federated learning framework, the global model update formula is as follows: Where M is the number of edge nodes. These are the local model parameters for the m-th node.

[0015] Secondly, the geological disaster early warning system based on the fusion analysis of hyperspectral and Internet of Things data provided in this application adopts the following technical solution: A geological disaster early warning system based on the fusion analysis of hyperspectral and IoT data includes: The data acquisition module is used to acquire hyperspectral remote sensing images and IoT sensor data; The preprocessing module is used to perform non-negative tensor ring decomposition and quantum entanglement optimization on hyperspectral data, and to construct mechanically driven dynamic topology for IoT data. The feature fusion module is used to concatenate spectral-spatial features and temporal features into a multidimensional tensor, and to perform cross-modal feature interaction through separable convolutional kernels. The optimization decision module is used to perform quantum-classical hybrid parameter optimization and to achieve edge-cloud collaborative early warning based on risk probability.

[0016] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention solves the limitations of traditional single data sources in terms of spatiotemporal resolution and physical correlation by using non-negative tensor ring decomposition of hyperspectral data and dynamic topology modeling of the Internet of Things. It can realize multi-dimensional joint extraction of spectral-spatial-mechanical features, and significantly enhance the characterization accuracy of the deformation evolution law of rock and soil. 2. This invention employs the quantum Ising model and the classical gradient sharing mechanism to transform the tensor decomposition parameter optimization problem into a global search problem solvable by quantum annealing. This overcomes the defect of traditional heuristic algorithms being prone to getting trapped in local optima and enables rapid convergence and stable generalization of model parameters in complex geological environments. 3. This invention embeds the constraints of the geotechnical equilibrium equation into the loss function, forcing the model prediction results to conform to the physical laws of stress-strain, avoiding mechanical contradictions that may be caused by purely data-driven methods, and has the interpretability of physical laws and engineering guidance value.

[0017] 4. This invention reduces computational load by using a binary quantization model at the edge, ensuring millisecond-level real-time anomaly detection; the cloud-based dynamic updating of physical constitutive equations based on federated learning enables the synergy of high-performance early warning and long-term model iterative optimization in resource-constrained scenarios.

[0018] 5. This invention constructs a dynamic adjacency matrix for sensors based on mechanical similarity and data confidence, which adapts to changes in formation stress and network fluctuations; combined with differential privacy noise injection in federated learning, it can balance model accuracy and privacy security in multi-node data collaboration. Attached Figure Description

[0019] Figure 1 This is the algorithm flowchart of this application; Detailed Implementation

[0020] The following is in conjunction with the appendix Figure 1 This application will be described in further detail below.

[0021] Example 1: A geological disaster early warning algorithm based on the fusion analysis of hyperspectral and IoT data, referring to... Figure 1 ,include: S1. Preprocess hyperspectral data and IoT data to extract spectral-spatial features and dynamic topological temporal features, respectively; The preprocessing step S1 first involves performing nonnegative tensor ring decomposition on the hyperspectral data to extract spectral-spatial features. Specifically, the input hyperspectral data cube is represented as a four-dimensional tensor. Where H is the spatial height, W is the spatial width, B is the number of spectral bands, and T is the time series length. To achieve efficient data compression and feature extraction, the Non-negative Tensor Chain Decomposition (NTCD) method is used to... It can be decomposed into a chained outer product of multiple core tensors, and its mathematical expression is: in Let R1, R2, ..., R be the k-th core tensor. N To decompose the rank parameter, R represents the tensor outer product operation. k The k-th rank parameter determines the complexity of tensor ring decomposition.

[0022] Preferably, the core tensor G (k) The nonnegativity constraint (G) must be satisfied. (k) ≥0) to preserve the physical meaning of spectral reflectance and avoid spectral distortion introduced by negative values.

[0023] To optimize the selection of rank parameters and improve decomposition accuracy, a quantum entanglement encoding mechanism is further introduced. Each rank parameter R... k Mapped to qubit q k The quantum state of is entangled with different rank parameters through a controlled NOT gate (CNOT) in a quantum circuit, and its quantum state expression is: Where q1 and q2 are qubit numbers, and |ψ> represents the entangled state. This design forces different rank parameters to adjust collaboratively during optimization, avoiding the local optima problem caused by traditional independent optimization. The quantum entanglement encoding result is iteratively updated using a quantum annealing algorithm until the tensor reconstruction error is reached. Converges to the set threshold.

[0024] Simultaneously, dynamic topology modeling is performed on IoT sensor data to extract temporal features. Sensor node data contains multi-dimensional parameters such as displacement, pore water pressure, and tilt angle, and its temporal sequence is denoted as... Where D represents the sensor parameter dimension and T represents the time step. To characterize the dynamic correlation characteristics of the sensor network, a mechanically driven adjacency matrix is ​​first constructed. The connection weight A between node i and node j is... ij The calculation expression is determined jointly by the soil shear force propagation model and data confidence level: Where F i =[τ i ,σ n,i ] is the shear force vector at node i, τ i For shear stress, σ n,i For normal stress, Confidence i To score the confidence level of the data, σ and α are learnable parameters. Shear force vector F i The results are obtained by calculating the node location coordinates and formation mechanical parameters (such as internal friction angle and cohesion), and preferably by using the Mohr-Coulomb criterion for iterative solution.

[0025] Based on the dynamic adjacency matrix, a dynamic graph neural network (DTGNN) is used to extract temporal features from the sensor. For the l-th layer of the network, the node features H (l) The updated formula is: H (l+1) =ReLU(AH (l) W (l) ); in Let be the feature matrix of the l-th layer nodes, N be the number of sensor nodes, and d be the feature dimension. W is a trainable weight matrix, ReLU is the activation function, and W is the weight matrix. (l) This is a trainable weight matrix used for feature space transformation. The adjacency matrix A is dynamically adjusted based on real-time sensor data, enabling it to adapt to changes in formation stress and network topology fluctuations.

[0026] Nonnegative tensor ring decomposition is used to eliminate redundancy in hyperspectral data while preserving physical interpretability; Quantum entanglement coding optimizes the selection of the rank parameter through quantum computing, thus overcoming the local optima defect of traditional heuristic methods. The mechanically driven adjacency matrix embeds the laws of geotechnical mechanics into topological modeling, enhancing the representation ability of dynamic correlations in sensor networks; the dynamic graph neural network achieves hierarchical extraction of temporal features through trainable weight matrices and dynamic adjacency matrices.

[0027] Step S1 achieves the following effects through the above technical solution: Dimensionality reduction and feature enhancement of hyperspectral data while preserving the multidimensional correlation between spectrum, space, and time; dynamic modeling of sensor network topology to adapt to changes in formation stress and fluctuations in data quality; and synergy between quantum optimization and classical computation to improve parameter solving efficiency and model generalization ability.

[0028] S2. Fuse the spectral-spatial features and temporal features to generate cross-modal joint features; The cross-modal feature fusion process in step S2 first involves fusing the hyperspectral tensor features extracted in step S1. With IoT time-series characteristics Perform dimensional alignment and concatenation. To achieve a unified representation of heterogeneous data, [the following steps are taken]. Expand along the parameter dimension Align its third dimension with the number of sensor nodes N, and then concatenate it with H along the fourth dimension to generate a four-dimensional joint tensor. This operation can preserve the multi-scale characteristics of spectral-spatial features and the dynamic evolution of temporal features.

[0029] To reduce the computational complexity of high-dimensional tensor operations, a lightweight separable tensor convolution kernel (LC-TensorNet) is further employed for cross-modal interaction. A three-dimensional separable convolution kernel is defined. in The convolution kernels operate on the spectral, spatial, and node dimensions respectively, with k1, k2, and k3 representing the kernel sizes. Preferably, each kernel performs a one-dimensional convolution operation independently, and the output features are calculated as follows: Where * denotes a one-dimensional convolution operation. Let k1 be the m-th convolutional kernel and k2 be the n-th convolutional kernel. k1, k2, and k3 are separable convolutional kernels that operate on the spectral, spatial, and node dimensions, respectively. The separable design can reduce the computational complexity of the original 3D convolution from O(R1R2Nk1k2k3) to O(R1k1+R2k2+Nk3), adapting to the real-time requirements of edge devices.

[0030] To enhance the physical interpretability of the model, geotechnical equilibrium constraints are embedded in the loss function. The stress tensor predicted by the model is defined as follows: The volume force vector is According to the principles of continuum mechanics, it must satisfy the equilibrium equations. σ t To predict the stress tensor, b t It is a volume force vector. The divergence operator is used to calculate the spatial rate of change of the stress tensor. Preferably, the divergence term is calculated using automatic differentiation techniques. The equilibrium condition is then added as a regularization term to the loss function: Joint optimization of data-driven loss With respect to physical constraints, the total loss function is: Where λ is the equilibrium coefficient, used to adjust the consistency weight between data fitting and physical laws. This design can force the model prediction results to conform to the static equilibrium conditions of soil and rock, avoiding mechanical contradictions that may arise from purely data-driven methods. Furthermore, the parameters of the fused features are optimized using a quantum-classical hybrid optimization algorithm. The tensor decomposition rank parameter selection problem is mapped to the quantum Ising model, whose Hamiltonian is defined as: H = -∑ i<j J ij q i q j -∑ i h i q i ; Where q i ∈{0,1} represents the activation state of the i-th rank parameter, J ij h is the coupling coefficient, describing the correlation strength between rank parameters. i This is a bias term, reflecting the impact of a single rank parameter on the reconstruction error. ij with h i The value is determined by the tensor reconstruction error. The second derivative is calculated. The quantum annealing result is used as the initial parameters of the classical neural network, and the quantum circuit parameters θ and the classical network weights W are jointly updated through a gradient sharing mechanism. The update formula is: This collaborative optimization mechanism can overcome the slow convergence speed of traditional alternating optimization methods, while leveraging the parallelism of quantum computing to accelerate the search for the global optimal solution.

[0031] Tensor splicing and alignment: Solving the heterogeneity problem between the multidimensionality of hyperspectral data and the dynamic nature of IoT time-series data, and establishing a unified feature space; Separable convolutional kernels: reduce computational complexity through dimensional decoupling, adapting to the real-time requirements of edge deployments; Physical constraint embedding: The equilibrium equations of geotechnical mechanics are used as prior knowledge, and the output of the constraint model conforms to physical laws. Quantum-classical co-optimization: Combining the global search capability of quantum computing with the local optimization efficiency of classical gradient descent to improve the accuracy of parameter solving.

[0032] Step S2 achieves the following effects through the above technical solution: The deep integration of hyperspectral and IoT data captures the nonlinear correlation between multimodal data; lightweight convolution operations ensure the real-time inference capabilities of edge devices; physical constraints enhance the credibility and interpretability of model prediction results; quantum-classical co-optimization accelerates model parameter convergence and avoids local optima.

[0033] S3. Optimize the parameters of the fused features using a quantum-classical hybrid optimization algorithm and calculate the disaster risk probability. The quantum-classical hybrid optimization process in step S3 first inputs the cross-modal fused features generated in step S2 into the quantum computing module, and optimizes the tensor decomposition rank parameters using the quantum Ising model. To achieve global optimal solution search, a Hamiltonian model is constructed to describe the correlation between the rank parameters, and its mathematical expression is: H = -∑ i<j J ij q i q j -∑ i h i q i ; Where q i ∈{0,1} represents the activation state of the i-th rank parameter, J ij The coupling coefficient reflects the rank parameter R. i With R j The interaction strength, h i This is a bias term, representing the effect of a single rank parameter on the tensor reconstruction error. Preferably, J... ij with h i The value is determined by the tensor reconstruction error. Calculation of the second and first derivatives: This design maps the tensor decomposition problem to a quantum annealing-solvable ground state search problem, overcoming the local optima limitation of classical optimization algorithms through the quantum tunneling effect. After solving for the Hamiltonian ground state using the quantum annealing algorithm, the optimal rank parameter combination is used as the initialization parameters for the classical neural network. To achieve co-optimization of quantum and classical computing, a gradient sharing mechanism is further established: the quantum circuit parameters θ (including entangled state configuration and quantum gate angle) are updated through the classical backpropagation algorithm, with the update formula as follows: Where η is the learning rate. This is a data-driven loss function. This mechanism enables the quantum parameters to be dynamically adjusted based on the model's prediction error, achieving end-to-end joint optimization of quantum computing and classical training.

[0034] To quantify the uncertainty of geological disaster risk, data noise and model errors are further encoded into quantum states. This is done by applying the core tensor G of the NTCD decomposition in step S1.(k) Perform singular value decomposition (SVD) to obtain the unitary matrix U. k and V k Mapping it to the unitary operator U of a quantum circuit quantum : Where θ k This is an adjustable phase parameter. The data uncertainty ∈ ~N(0,σ) 2 Encoded as quantum phase noise, input quantum state ρ in Entanglement operation with noise state |∈>: The final risk probability is measured by the projective operator M = |ψ fail ><ψ fail Calculation: P = Tr(ρ) out ·M); Where |ψ fail > represents the quantum eigenstate of a disaster. This process propagates data uncertainty to the risk probability output, providing a probabilistic assessment basis for early warning decisions.

[0035] The quantum Ising model transforms the rank selection problem of tensor decomposition into a quantum optimization problem, and uses quantum parallelism to accelerate the search for the global optimal solution. Gradient sharing mechanism: Establish a joint optimization path between quantum parameters and classical networks to avoid error accumulation caused by traditional step-by-step optimization; Uncertainty quantum encoding: Through phase noise injection and projection measurement, end-to-end propagation of data noise to risk probability is achieved; SVD-quantum circuit mapping: The mathematical structure of tensor decomposition is embedded into the quantum computing process, preserving physical interpretability. Step S3 achieves the following effects through the above technical solution: Quantum annealing optimization breaks through the local optima limitation of classical algorithms and improves the accuracy of tensor decomposition; quantum-classical gradient sharing enables end-to-end parameter collaborative updates and accelerates model convergence; data uncertainty is propagated to early warning output through quantum state entanglement, enhancing the reliability of decision-making.

[0036] S4. Based on the optimization results and risk probability, execute edge-cloud collaborative early warning decisions.

[0037] Step S4, the edge-cloud collaborative early warning decision-making process, first deploys a binary quantization model at the edge to achieve real-time anomaly detection. The weight parameters of the lightweight tensor convolutional network (LC-TensorNet) optimized in step S3 are then... Quantization is performed to 2-bit precision, and its weighted quantization formula is defined as follows: Wbin =Δ·sign(W), Where n is the total number of weight parameters, Δ is the dynamic scaling factor, sign(·) is the sign function, and W bin This is the quantized 2-bit weight matrix. Preferably, the quantized model is implemented using fixed-point arithmetic, and the inference latency satisfies t. infer ≤20ms, adapting to the resource constraints of edge devices. This design reduces memory footprint and energy consumption by lowering computational precision, while preserving the model's sensitivity to geological hazard characteristics.

[0038] The cloud-based system dynamically updates the constitutive equations of geotechnical mechanics based on real-time sensor data to improve long-term prediction accuracy. The stress-strain relationship of the soil mass is defined as a nonlinear function. in For the current stress tensor, Let f be the displacement gradient, Θ be the formation mechanical parameters (such as elastic modulus and cohesion), and f be the displacement gradient. MLP For multilayer perceptrons, model the stress-strain nonlinear relationship. The displacement gradient is calculated by differential analysis of sensor displacement data. A multilayer perceptron (MLP) model is used to model the nonlinear relationship. The network input layer dimension is 3x3+3=12, the hidden layer activation function is LeakyReLU, and the output layer uses linear activation. Preferably, the formation parameter Θ is updated by assimilating it with real-time monitoring data using a Bayesian filtering algorithm to ensure consistency between the physical model and the field conditions.

[0039] Furthermore, a federated learning framework is used to aggregate parameters of multiple edge nodes in the cloud. Let W be the local model parameters of the M edge nodes. edge,1 ,…,W edge,M The formula for updating the global model in the cloud is: Where k is the communication round index, W cloud W represents the global model parameters in the cloud. edge,m Let W be the local model parameters for the m-th edge node. Preferably, differential privacy technology is used during the aggregation process to add Gaussian noise W to the local parameters. edge,m ←W edge,m +N(0,σ 2 I) To protect data privacy, σ is the standard deviation of Gaussian noise, controlling the strength of privacy protection. The updated global model is distributed to edge nodes via wireless communication, forming a closed-loop feedback of "real-time edge response - long-term cloud optimization".

[0040] Binary quantization model: By reducing computational precision, it meets the real-time requirements of edge computing and works in synergy with the lightweight tensor convolution design in step S2; Mechanical constitutive equation update: Combine the physical constraint loss function in step S2 to achieve bidirectional coupling between the data-driven model and physical laws; Federation parameter aggregation: Based on the quantum-classical hybrid optimization results of step S3, the generalization ability of the model is improved through distributed learning; Privacy protection mechanism: The confidence weighting mechanism of step S1 is added to the data sharing level to form an end-to-end security early warning system.

[0041] Step S4 achieves the following effects through the above technical solution: Edge-based quantization models ensure real-time performance and low power consumption in geological hazard anomaly detection; cloud-based dynamic updates of mechanical constitutive equations improve the physical consistency of long-term predictions; and a federated learning framework achieves the dual goals of multi-node collaborative data training and privacy protection. Example 2: A geological disaster early warning system based on the fusion analysis of hyperspectral and Internet of Things data, comprising: The data acquisition module is used to acquire hyperspectral remote sensing images and IoT sensor data; Hyperspectral remote sensing images (spectral resolution better than 10nm, spatial resolution ≤1m) are acquired through satellite or drone platforms, and data from IoT sensors deployed in the monitoring area (including multi-dimensional parameters such as displacement, pore water pressure, tilt angle, and shear stress) are received simultaneously.

[0042] Heterogeneous data compatibility: Supports multi-source heterogeneous data formats (such as hyperspectral ENVI format and sensor JSON protocol), and has a built-in clock synchronization mechanism to ensure spatiotemporal data alignment.

[0043] The preprocessing module is used to perform non-negative tensor ring decomposition and quantum entanglement optimization on hyperspectral data, and to construct mechanically driven dynamic topology for IoT data. Hyperspectral data compression: The hyperspectral cube was decomposed using nonnegative tensor ring decomposition (NTCD). Perform dimensionality reduction. The decomposition formula is: Among them G (k) ≥0 represents the core tensor, R k Let be the rank parameter, and use quantum entanglement encoding to force nonlinear correlations between different rank parameters.

[0044] Internet of Things (IoT) data modeling: Construct a dynamic adjacency matrix A driven by mechanics, with the weights calculated as follows: Where F i For the nodal shear vector, Confidence i For confidence scores, σ and α are learnable parameters.

[0045] The feature fusion module is used to concatenate spectral-spatial features and temporal features into a multidimensional tensor, and to perform cross-modal feature interaction through separable convolutional kernels. Multidimensional tensor splicing: combining the spectral and spatial features output by NTCD Temporal features extracted by dynamic graph neural networks Concatenate along the node dimensions to form a joint tensor

[0046] Cross-modal interaction: through separable convolutional kernels Feature fusion is performed, and the calculation formula is as follows: K1, K2, and K3 act on the spectral, spatial, and nodal dimensions, respectively.

[0047] The optimization decision module is used to perform quantum-classical hybrid parameter optimization and to achieve edge-cloud collaborative early warning based on risk probability.

[0048] Quantum-classical hybrid optimization: Mapping the tensor decomposition problem to the quantum Ising model, the Hamiltonian is defined as: H = -∑ i<j J ij q i q j -∑ i h i q i ; Where q i ∈{0,1} represents the activation state of the rank parameter, J ij with h i The parameters are calculated from the second derivative of the reconstruction error. The quantum annealing results are used as initial parameters for the classical network, and then jointly optimized through a gradient sharing mechanism.

[0049] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Identical components are represented by the same reference numerals. Therefore, all equivalent changes made to the structure, shape, and principle of this application should be covered within the scope of protection of this application.

Claims

1. A geological disaster early warning algorithm based on the fusion analysis of hyperspectral and Internet of Things data, characterized in that, include: S1. Preprocess hyperspectral data and IoT data to extract spectral-spatial features and dynamic topological temporal features, respectively; S2. Fuse the spectral-spatial features and temporal features to generate cross-modal joint features; S3. Optimize the parameters of the fused features using a quantum-classical hybrid optimization algorithm, and calculate the probability of disaster risk; S4. Based on the optimization results and risk probability, execute edge-cloud collaborative early warning decisions.

2. The geological disaster early warning algorithm based on hyperspectral and IoT data fusion analysis according to claim 1, characterized in that, The preprocessing of hyperspectral data in step S1 includes: Where χ is the input hyperspectral cube, Let R1, ..., R be the core tensor after decomposition. N It is a rank parameter; The nonlinear correlation between rank parameters is established through quantum entanglement encoding, specifically mapped as follows: Where q1 and q2 are qubits, and |ψ> represents an entangled state.

3. The geological disaster early warning algorithm based on the fusion analysis of hyperspectral and IoT data as described in claim 1, characterized in that, The preprocessing of IoT data in step S1 includes: The dynamic adjacency matrix of sensor nodes is constructed, and its weight calculation expression is as follows: Where F i Let i be the shear vector at node i. i Let σ and α be the data confidence level, and α be the learnable parameters. Temporal features are extracted using a dynamic graph neural network, and its update formula is as follows: A (l+1) =ReLU(AH (l) W (l) ); Where H (l) For the features of the l-th layer nodes, W (l) This is a trainable weight matrix.

4. The geological disaster early warning algorithm based on hyperspectral and IoT data fusion analysis according to claim 1, characterized in that, The cross-modal joint feature fusion in step S2 includes: Hyperspectral tensor features and IoT temporal features are concatenated along the parameter dimension to form a multidimensional tensor; Separable tensor convolution kernels are used for cross-modal feature interaction computation.

5. The geological disaster early warning algorithm based on hyperspectral and IoT data fusion analysis according to claim 1, characterized in that, The fusion process also includes: The constraint of the geotechnical equilibrium equation is embedded in the loss function, and its expression is as follows: Where σ t To predict the stress tensor, b t It is a volume force vector. • is the divergence operator.

6. The geological disaster early warning algorithm based on hyperspectral and IoT data fusion analysis according to claim 1, characterized in that, The quantum-classical hybrid optimization in step S3 includes: Mapping the tensor decomposition problem to the quantum Ising model, its Hamiltonian expression is: H=-∑ i<j J ij q i q j -∑ i h i q i ; Where, q i ∈{0,1} represents the rank parameter selection state, J ij with h i Calculated from tensor reconstruction error; The quantum annealing results are used as initialization parameters for a classical neural network, and jointly optimized through a gradient sharing mechanism.

7. The geological disaster early warning algorithm based on hyperspectral and IoT data fusion analysis according to claim 1, characterized in that, The quantum optimization also includes: Encoding data uncertainty as quantum phase noise, its quantum state expression is as follows: Where ∈ represents data uncertainty, U quantum For quantum circuit unitary operators.

8. The geological disaster early warning algorithm based on hyperspectral and IoT data fusion analysis according to claim 1, characterized in that, The edge-cloud collaborative early warning decision-making includes: A binary quantization model is deployed at the edge for real-time anomaly detection, and its weight quantization formula is as follows: Where n is the total number of weights, and Δ is the dynamic scaling factor; The geotechnical constitutive equations are dynamically updated in the cloud based on real-time data.

9. The geological disaster early warning algorithm based on hyperspectral and IoT data fusion analysis according to claim 8, characterized in that, The cloud update also includes: By aggregating parameters from multiple edge nodes using a federated learning framework, the global model update formula is as follows: Where M is the number of edge nodes. These are the local model parameters for the m-th node.

10. A geological disaster early warning system based on the fusion analysis of hyperspectral and Internet of Things (IoT) data, comprising the geological disaster early warning algorithm based on the fusion analysis of hyperspectral and IoT data according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire hyperspectral remote sensing images and IoT sensor data; The preprocessing module is used to perform non-negative tensor ring decomposition and quantum entanglement optimization on hyperspectral data, and to construct mechanically driven dynamic topology for IoT data. The feature fusion module is used to concatenate spectral-spatial features and temporal features into a multidimensional tensor, and to perform cross-modal feature interaction through separable convolutional kernels; The optimization decision module is used to perform quantum-classical hybrid parameter optimization and to achieve edge-cloud collaborative early warning based on risk probability.

Images