Mine disaster monitoring data analysis and early warning method, device, equipment and medium

Abstract

The invention relates to a mine disaster monitoring data analysis and early warning method and device, equipment and a medium. The method comprises the following steps: acquiring multi-source sensing data and a three-dimensional geologic model of a mine; based on a PSO-CNN-LSTM hybrid algorithm, performing time sequence prediction and feature extraction on the multi-source sensing data and the three-dimensional geologic model to obtain disaster classification levels and fusion feature data; when the disaster classification level meets a preset requirement, performing feature masking on the fused feature data to obtain local model data; and based on a federated learning method, according to a security aggregation protocol, performing global fusion on the local model data to obtain shared early warning data. By adopting the method, mine sensing data sharing can be realized for training a comprehensive early warning model, and the accuracy of disaster prediction is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of mine safety management, and in particular relates to a mine disaster monitoring data analysis and early warning method, device, equipment and medium. Background Art

[0002] With the development of intelligent technology, mine safety management is gradually shifting from manual inspections to automation and data-driven management. Disaster early warning models are built through data collection and analysis. Based on real-time monitoring of mine data and the disaster early warning models, mine safety ratings are generated.

[0003] Existing intelligent mine safety management technology collects multi-parameter data such as electromagnetic, microseismic, acoustic emission, and dynamic gas outburst in the mine, performs multimodal data fusion and feature extraction, and constructs a weight model based on the critical values of sensitive parameters of various types of data. Based on the comprehensive early warning model, disaster prediction is achieved and the reliability of early warning is improved.

[0004] However, the current intelligent mine safety management technology has the problem of data silos. The comprehensive early warning model is trained only through a single data source. In addition, the geological data of each mining area involves commercial secrets or sensitive information, making it impossible to share data for training the comprehensive early warning model, which reduces the accuracy of disaster prediction. Summary of the Invention

[0005] Based on this, it is necessary to provide a mine disaster monitoring data analysis and early warning method, device, equipment and medium that can realize training data sharing and improve the accuracy of disaster model prediction in response to the above technical problems.

[0006] In a first aspect, the present application provides a mine disaster monitoring data analysis and early warning method, comprising:

[0007] Acquire multi-source sensor data and 3D geological models of mines;

[0008] Based on the PSO-CNN-LSTM hybrid algorithm, time series prediction and feature extraction are performed on multi-source sensor data and 3D geological models to obtain disaster classification levels and fused feature data;

[0009] When the disaster classification level meets the preset requirements, the fused feature data is masked to obtain the local model data;

[0010] Based on the federated learning method and according to the security aggregation protocol, local model data are globally fused to obtain shared warning data.

[0011] Furthermore, based on the PSO-CNN-LSTM hybrid algorithm, time series prediction and feature extraction are performed on multi-source sensor data and 3D geological models to obtain disaster classification levels and fused feature data, including:

[0012] The following formula is used to calculate the high-frequency sensor data:

[0013]

[0014] in, is the high-frequency sensor data, DWT is the discrete wavelet transform function, z k It is to observe multi-source sensor data;

[0015] Calculate the prediction covariance based on the high-frequency sensor data;

[0016] Denoised sensor data is obtained based on the removed high-frequency sensor data and the predicted covariance;

[0017] The denoised sensor data is calculated using the following formula:

[0018]

[0019] in, is the denoised sensor data at the current moment, is the denoised sensor data at the previous moment, K k is the Kalman gain, H k is the observation matrix, It is to remove high-frequency sensor data;

[0020] Segment the three-dimensional geological model to obtain a geological model region set;

[0021] Based on the PSO-CNN-LSTM hybrid algorithm, the temporal characteristics of the denoised sensor data and the spatial characteristics of the geological model area set are fused and analyzed to obtain the disaster classification level and fused feature data.

[0022] Furthermore, based on the PSO-CNN-LSTM hybrid algorithm, the temporal characteristics of the denoised sensor data and the spatial characteristics of the geological model area set are fused and analyzed to obtain the disaster classification level and fused feature data, including:

[0023] Obtaining denoised sensor data and a set of geological model regions;

[0024] Based on the LSTM algorithm, the candidate information state is obtained by constructing the input gate;

[0025] The candidate information state is obtained by the following formula:

[0026] C t =tanh(W C ·[h t-1 ,x t ]+b C )

[0027] Among them, Ct is the candidate information state, W C is the input gate weight matrix, h t-1 is the hidden data of the previous moment, x t is the current input denoised sensor data, b c is the input bias term;

[0028] According to the candidate information state, the cell state information is updated to obtain the sensing time series data;

[0029] Among them, the sensing time series data is obtained through the following formula:

[0030] h t =o t ⊙tanh(C t )

[0031] Among them, h t is the sensor time series data, o t is the output hidden state data, ⊙ is the element-wise multiplication, C t It is to update information data;

[0032] Among them, the output hidden state data is obtained by the following formula:

[0033]

[0034] Among them, t is the output hidden state data, w0 is the output gate weight matrix, σ is the activation function, h t-1 is the hidden data of the previous moment, x t is the current input denoised sensor data, b0 is the output gate bias term;

[0035] Based on the CNN algorithm, the spatial characteristics of the geological model area set are extracted to obtain the geological characteristic map. The sensor time series data and the geological characteristic map are then comprehensively analyzed to obtain the disaster classification level and fusion feature data.

[0036] The calculation formula for disaster classification level is:

[0037] p=Softmax(W cls f fusion +b cls )

[0038] Where p is the disaster classification level, Softmax is the classification function, W cls is the classification weight, f fusion is the fusion feature data, b cls is the bias term;

[0039] Based on the PSO algorithm, the cross entropy loss is minimized and the convolution kernel parameters and LSTM hidden data dimension parameters are optimized.

[0040] Furthermore, when the disaster classification level meets the preset requirements, the fused feature data is masked to obtain local model data, including:

[0041] Use the following formula to perform feature masking on the fused sensor data to obtain masked sensor data:

[0042] X clean =X global ⊙M mask

[0043] Among them, X clean is to mask the sensor data, X global is the fused sensor data, ⊙ is the bitwise AND operator, M mask is the feature mask matrix;

[0044] According to the masked sensing data, gradient sensing data is obtained based on the time series analysis algorithm;

[0045] Performing gradient clipping on the gradient sensing data to obtain clipped sensing data;

[0046] Noise is added to the cropped sensor data to obtain local model data.

[0047] Furthermore, the gradient sensing data is subjected to gradient clipping to obtain clipped sensing data, including:

[0048] Use the following formula to perform gradient clipping on the gradient sensor data to obtain clipped sensor data:

[0049]

[0050] in, It is the cropping sensor data, Top is the interception function, is the gradient sensing data, and S is the interception ratio.

[0051] Furthermore, the cropped sensor data is noise-added to obtain local model data, including:

[0052] Use the following formula to add noise and obtain local model data:

[0053]

[0054] Among them, H n is the local model data, is the cropped sensor data, Laplace is the Laplace distribution function, Δ is the feature sensitivity, and ∈ is the noise threshold.

[0055] Furthermore, based on the federated learning method and the secure aggregation protocol, local model data is globally integrated to obtain shared warning data, including:

[0056] According to the local model data, the global model data is obtained based on the federated learning method;

[0057] Among them, the global model data is obtained using the following formula:

[0058]

[0059] Among them, G g is the global model data, N is the number of mines participating in federated learning, and H i is the i-th local model data;

[0060] According to the security aggregation protocol, the global model data is globally integrated to obtain shared warning data;

[0061] The shared warning data is obtained using the following formula:

[0062] E(G g )=Ec(R,G g )

[0063] Among them, E(G g ) is the shared warning data, Ec is the encryption function, R is the public key, G g It is the global model data.

[0064] In a second aspect, the present application also provides a mine disaster monitoring data analysis and early warning device, comprising:

[0065] Information acquisition module, used to obtain multi-source sensor data and three-dimensional geological models of the mine;

[0066] The level determination module is used to perform time series prediction and feature extraction on multi-source sensor data and 3D geological models based on the PSO-CNN-LSTM hybrid algorithm to obtain disaster classification levels and fused feature data;

[0067] The information processing module is used to mask the fused feature data and obtain local model data when the disaster classification level meets the preset requirements;

[0068] The information sharing module is used to globally fuse local model data based on the federated learning method and the security aggregation protocol to obtain shared warning data.

[0069] In a third aspect, the present application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements any mine disaster monitoring data analysis and early warning method described in the first aspect of the present application.

[0070] In a fourth aspect, the present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements any mine disaster monitoring data analysis and early warning method described in the first aspect of the present application.

[0071] The above-mentioned mine disaster monitoring data analysis and early warning method, device, equipment and medium obtain multi-source sensor data and three-dimensional geological model of the mine; based on the PSO-CNN-LSTM hybrid algorithm, time series prediction and feature extraction of the multi-source sensor data and three-dimensional geological model are performed to obtain disaster classification level and fused feature data; when the disaster classification level meets the preset requirements, the fused feature data is feature masked to obtain local model data; based on the federated learning method, according to the security aggregation protocol, the local model data is globally fused to obtain shared early warning data, thereby realizing the integration and sharing of mine disaster monitoring data, increasing the training data sources of the early warning models of each mining area, and improving the accuracy of disaster prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0073] Figure 1 A flowchart of a mine disaster monitoring data analysis and early warning method provided as an exemplary embodiment of the present application;

[0074] Figure 2 A flowchart of a method for intelligent management and control of mine disaster risks provided as an exemplary embodiment of this application;

[0075] Figure 3 A structural block diagram of a mine disaster monitoring data analysis and early warning device provided as an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0076] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0077] The mine disaster monitoring data analysis and early warning method, device, equipment and medium provided in the embodiments of the present application can be applied to application scenarios in which global shared disaster early warning data is generated by acquiring multi-source sensor data in mines.

[0078] In one embodiment, Figure 1 As shown, a mine disaster monitoring data analysis and early warning method is provided. This embodiment uses the method applied to a terminal as an example for illustration. It is understandable that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0079] Step S101: Acquire multi-source sensor data and a three-dimensional geological model of a mine.

[0080] Specifically, the terminal acquires sensor data from each sensor within the mine and constructs a comprehensive 3D geological model of the mine. Illustratively, the sensor data includes, but is not limited to, gas concentration data, temperature data, vibration data, and acoustic wave data. Optionally, the 3D geological model can be a millimeter-scale 3D geological model constructed by combining drone oblique photography and lidar scanning. This 3D geological model includes, but is not limited to, the geological features of the mine.

[0081] Step S102: Based on the PSO-CNN-LSTM hybrid algorithm, time series prediction and feature extraction are performed on the multi-source sensor data and the three-dimensional geological model to obtain the disaster classification level and fusion feature data.

[0082] Specifically, the terminal preprocesses the acquired multi-source sensor data and 3D geological model, producing noise-removed sensor data and regional geological model data. Using a PSO-CNN-LSTM hybrid algorithm, the terminal performs time series prediction on the preprocessed multi-source sensor data to obtain time series features. Feature extraction is performed on the 3D geological model to obtain spatial features. These temporal and spatial features are then fused and classified probabilities are calculated to produce fused feature data and a hazard classification level. Schematically, preprocessing includes noise removal and regional segmentation.

[0083] Step S103: When the disaster classification level meets the preset requirements, feature masking is performed on the fused feature data to obtain local model data.

[0084] Specifically, when the disaster classification level meets preset requirements, feature masking is performed on the fused feature data to generate local model data. Illustratively, the preset requirements can be set based on the user's disaster response needs. Optionally, feature masking can include information masking, information clipping, and noise addition.

[0085] Step S104: Based on the federated learning method and in accordance with the security aggregation protocol, the local model data is globally integrated to obtain shared warning data.

[0086] Specifically, the terminal integrates global and local model data based on a federated learning approach, performs global integration and adds security protocols based on a security aggregation protocol, and obtains shared warning data. In principle, the security protocol can be selected and edited based on the user's data security needs.

[0087] This embodiment obtains multi-source sensor data and a three-dimensional geological model of the mine, extracts and fuses features of the data and model based on a PSO-CNN-LSTM hybrid algorithm, obtains disaster classification levels and fused feature data, and performs feature masking and security processing on the fused feature data according to the disaster classification levels. Based on a federated learning method and a secure aggregation protocol, shared early warning data is generated, and shareable training data is provided for model training. This increases the training data sources for early warning models in each mining area and improves the accuracy of disaster prediction.

[0088] In one embodiment, based on the PSO-CNN-LSTM hybrid algorithm, time series prediction and feature extraction are performed on multi-source sensor data and a three-dimensional geological model to obtain disaster classification levels and fused feature data, including:

[0089] S201, wherein the high-frequency sensing data is calculated using the following formula:

[0090]

[0091] in, is the high-frequency sensor data, DWT is the discrete wavelet transform function, z k It is to observe multi-source sensor data.

[0092] Specifically, based on the acquired multi-source sensor data, the high-frequency noise-removed high-frequency sensor data is calculated by using the formula for removing high-frequency noise. is the multi-source sensor data with high-frequency noise removed. The high-frequency noise may be the noise caused by electromagnetic interference of the device. Optionally, the discrete wavelet transform function DWT decomposes the signal into high-frequency (detail coefficient) and low-frequency (approximate coefficient) components, and performs threshold processing on the high-frequency coefficient to remove the high-frequency noise and retain only the low-frequency part. Schematically, the multi-source sensor data z is observed. k is the input multi-source sensor data.

[0093] S202: Calculate the prediction covariance based on the high-frequency-removed sensing data.

[0094] Specifically, based on the high-frequency sensor data, the dynamic correlation and uncertainty of different multi-source sensor signals are quantified using the covariance matrix to obtain the predicted covariance. In principle, the predicted covariance can be used to correct the noise removal prediction error and improve the reliability of the noise removal.

[0095] S203, obtaining denoised sensor data based on the high-frequency sensor data and the predicted covariance;

[0096] The denoised sensor data is calculated using the following formula:

[0097]

[0098] in, is the denoised sensor data at the current moment, is the denoised sensor data at the previous moment, K k is the Kalman gain, H k is the observation matrix, It is to remove high-frequency sensor data.

[0099] Specifically, the denoised sensor data is obtained based on the high-frequency sensor data and the predicted covariance. is the denoised sensor data at the current moment. Optionally, the denoised sensor data at the previous moment is the denoised sensor data of the previous moment. Schematically, the Kalman gain K k is calculated by predicting the covariance. Alternatively, the observation matrix H k Is the linear transformation matrix that maps the state space to the observation space, which can be a constant matrix. Schematically, to high-frequency sensor data It is multi-source sensor data with high-frequency noise removed.

[0100] S204: Segment the three-dimensional geological model to obtain a geological model region set.

[0101] Specifically, the 3D geological model is segmented to obtain a set of geological model regions. Schematically, the 3D geological model can be segmented into multiple independent geological region data starting from an initial seed point using a region growing algorithm. The initial seed point can be set according to the user's requirements for mine region management.

[0102] S205, based on the PSO-CNN-LSTM hybrid algorithm, the temporal characteristics of the denoised sensor data and the spatial characteristics of the geological model area set are fused and analyzed to obtain the disaster classification level and fused feature data.

[0103] Specifically, a hybrid PSO-CNN-LSTM algorithm is used to perform a multimodal fusion analysis of the temporal features of the denoised sensor data and the spatial features of the 3D geological model region set, generating disaster classification levels and fused feature data. Schematically, the LSTM algorithm can be used to extract the long-term temporal features of the denoised sensor data through a gating mechanism. Alternatively, a CNN algorithm can be used to extract the spatial features of the 3D geological model region set. These spatial features are then fused and combined with the long-term temporal features to generate fused feature data. Based on this fused feature data, a disaster classification level is output through classification. Schematically, the PSO algorithm can be used to adjust and optimize the parameters of the CNN and LSTM.

[0104] This embodiment improves the reliability of multi-source sensor data by suppressing high-frequency noise and combining Kalman filtering to dynamically optimize sensor data, segments the three-dimensional geological model, and uses the PSO-CNN-LSTM hybrid algorithm to fuse the temporal characteristics of multi-source sensor data and the spatial geological characteristics of the mine to obtain more accurate disaster classification levels and fused feature data.

[0105] In one embodiment, based on the PSO-CNN-LSTM hybrid algorithm, the temporal characteristics of the denoised sensor data and the spatial characteristics of the geological model region set are fused and analyzed to obtain the disaster classification level and fused feature data, including:

[0106] S301, obtaining denoised sensor data and a geological model region set.

[0107] Specifically, denoised sensor data from which noise is removed and a set of segmented geological model regions are obtained.

[0108] S302, based on the LSTM algorithm, obtains the candidate information state by constructing an input gate;

[0109] The candidate information state is obtained by the following formula:

[0110] C t =tanh(W C ·[h t-1 ,x t ]+b C )

[0111] Among them, C t is the candidate information state, W C is the input gate weight matrix, h t-1 is the hidden data of the previous moment, x t is the current input denoised sensor data, b c is the input bias term.

[0112] Specifically, based on the LSTM algorithm, candidate information states are generated through the input gate to capture the potential temporal features of the current moment. Schematically, the candidate information state C t Is the candidate information state at the current moment, providing candidate feature information for subsequent cell state updates. Optionally, the input gate weight matrix W C It is used to fuse the historical hidden state with the current input information. Schematically, the hidden data h at the previous moment t-1 Is the hidden state data that carries historical time series information. The hidden state dimension can be calculated by optimizing the PSO algorithm. Optionally, the current input denoised sensor data x t It can be denoised sensor data. Schematically, the input bias term b c Used to adjust the baseline value of the candidate state, the bias term dimension can be obtained by optimizing the PSO algorithm.

[0113] S303, updating the cell state information according to the candidate information state to obtain sensing time series data;

[0114] Among them, the sensing time series data is obtained through the following formula:

[0115] h t =o t ⊙tanh(C t )

[0116] Among them, h t is the sensor time series data, o t is the output hidden state data, ⊙ is the element-wise multiplication, C t It is to update information data;

[0117] Among them, the output hidden state data is obtained by the following formula:

[0118]

[0119] Among them, t is the output hidden state data, w0 is the output gate weight matrix, σ is the activation function, h t-1 is the hidden data of the previous moment, x t is the current input denoised sensor data, and b0 is the output gate bias term.

[0120] Specifically, by multiplying the output hidden state data and the updated information data element by element, the current moment information and the hidden long-term time series features are proportionally filtered to obtain the output sensor time series data. t It is a feature matrix that combines long-term time features with current moment features. Optionally, the output hidden state data o tIs the output hidden state data. Schematically, the update information data C t is the candidate information state at the current moment, providing candidate feature information for subsequent cell state updates. Optionally, the output gate weight matrix w0 is used to fuse the historical hidden state with the current input information. Schematically, the activation function σ can be a Sigmoid function. Optionally, the hidden data h at the previous moment t-1 The hidden state data carrying historical time series information can be obtained by optimizing the PSO algorithm. t Optionally, the output gate bias term b0 is used to adjust the reference value of the gate signal, and the dimension of the bias term can be obtained by optimizing the PSO algorithm.

[0121] S304, based on the CNN algorithm, extracting the spatial features of the geological model area set to obtain a geological feature map, and comprehensively analyzing the sensor time series data and the geological feature map to obtain a disaster classification level and fusion feature data;

[0122] The calculation formula for disaster classification level is:

[0123] p=Softmax(W cls f fusion +b cls )

[0124] Where p is the disaster classification level, Softmax is the classification function, W cls is the classification weight, f fusion is the fusion feature data, b cls is the bias term.

[0125] Specifically, based on the CNN algorithm, the spatial features of the geological model area set are extracted by setting a convolution layer to generate a geological feature map. By setting a fully connected layer and a classification function, the geological feature map is cascaded with the sensor time series data output by the LSTM to obtain the disaster classification level and the fusion feature data. Schematically, the disaster classification level p is the disaster classification level obtained based on the multi-source sensor data and three-dimensional geological model classification of the mine, and can be a level set according to the user's disaster prevention and control needs. Optionally, the classification function Softmax is a normalized exponential function that maps the linear score to a probability distribution. Schematically, the classification weight W cls is the classification weight representing the spatial-temporal characteristics. The dimension of this matrix is determined by the fusion feature data and the number of disaster classifications. Optionally, the fusion feature data f fusion It is obtained by fusing sensor time series data and geological feature maps. Schematically, the bias term b cls It is used to adjust the classification decision boundary.

[0126] S305, based on the PSO algorithm, calculate the minimum cross entropy loss and optimize the convolution kernel parameters and LSTM hidden data dimension parameters.

[0127] Specifically, based on the PSO algorithm, each particle represents a set of hyperparameter combinations (including the number of convolution kernels and hidden data dimension parameters). By iteratively updating the particle position, the minimum loss solution is found to obtain the optimized convolution kernel parameters and LSTM hidden data dimension parameters.

[0128] This embodiment dynamically acquires time series features through the input and output gates of LSTM to obtain sensor time series data, combines CNN to extract the spatial features of the three-dimensional geological model, and uses multimodal feature fusion and Softmax classification to fuse the geological feature map and sensor time series data and classify the results. Particle swarm optimization (PSO) is used to adaptively adjust the number of convolution kernels and the number of LSTM hidden units to obtain disaster classification levels and fused feature data, realizing efficient collaborative modeling of the spatiotemporal characteristics of mine disasters, significantly improving classification accuracy, and enhancing the model's robustness to noise interference and adaptability to complex geological environments.

[0129] In one embodiment, when the disaster classification level meets the preset requirements, the fused feature data is masked to obtain local model data, including:

[0130] S401, using the following formula, perform feature masking on the fused sensor data to obtain masked sensor data:

[0131] X clean =X global ⊙M mask

[0132] Among them, X clean is to mask the sensor data, X global is the fused sensor data, ⊙ is the bitwise AND operator, M mask is the feature mask matrix.

[0133] Specifically, based on the coordinates of the mine’s sensitive information contained in the fused sensor data, a feature mask matrix with the same dimensions as the fused sensor data is set and located with the fused sensor data to remove the corresponding sensitive information. Schematically, the assessment rules for sensitive information can be set based on the user’s information protection requirements and may include sensor locations, sensor parameters, and geological details. Optionally, masking the sensor data X clean It is the feature data that removes sensitive information and retains the disaster-related features in the fused sensor data. global The output of the hybrid algorithm based on PSO-CNN-LSTM is the fusion sensor data obtained by fusion sensor time series data and geological feature map. Optionally, the feature mask matrix Mmask It is a matrix data with the same dimension as the fused sensor data, set according to the coordinate position of the sensitive information in the fused sensor data. It can be a binary matrix, which marks the coordinate position of the feature information to be retained by setting element 1, and marks the coordinate position of the sensitive information by setting element 0.

[0134] S402 , obtaining gradient sensing data based on the mask sensing data and a time series analysis algorithm.

[0135] Specifically, based on the time series analysis algorithm, the gradient of the masked sensor data at the current moment and the masked sensor data at adjacent time steps is calculated through the difference formula to obtain the gradient sensor data.

[0136] S403 , performing gradient clipping on the gradient sensing data to obtain clipped sensing data.

[0137] Specifically, the gradient sensing data is clipped according to the user's needs to obtain clipped sensing data. Schematically, gradient clipping is to constrain the gradient amplitude to avoid generating parameters that are too large.

[0138] S404: Add noise to the cropped sensor data to obtain local model data.

[0139] Specifically, in order to mask sensitive gradient information and prevent the original data from being inferred through reverse engineering, the terminal adds Laplace noise to the cropped sensor data to obtain local model data.

[0140] This embodiment selectively shields sensitive information and retains disaster-related features through feature masks, extracts dynamic changes in time series based on differential gradient calculation, generates data amplitude in combination with gradient clipping constraints, and adds Laplace noise to achieve differential privacy protection, ultimately generating local model data that balances privacy and effectiveness.

[0141] In one embodiment, performing gradient clipping on the gradient sensing data to obtain clipped sensing data includes:

[0142] Use the following formula to perform gradient clipping on the gradient sensor data to obtain clipped sensor data:

[0143]

[0144] in, It is the cropping sensor data, Top is the interception function, is the gradient sensing data, and S is the interception ratio.

[0145] Specifically, the gradient sensor data is clipped using the clipping function to obtain clipped sensor data. is the data after gradient clipping. Optionally, the clipping function Top retains the input gradient data according to the input clipping ratio. Schematically, the gradient sensing data The gradient sensor data is obtained based on the time series analysis algorithm. Schematically, the interception ratio S controls the intensity of the retained gradient and can be set according to the user's requirements for communication overhead and shared sensor data.

[0146] This embodiment selectively retains the original gradient through the interception function to obtain the tailored sensor data that meets the user's needs, which can reduce the amount of communication transmission data and improve the stability of shared data.

[0147] In one embodiment, the cropped sensor data is subjected to noise addition to obtain local model data, including:

[0148] Use the following formula to add noise and obtain local model data:

[0149]

[0150] Among them, H n is the local model data, is the cropped sensor data, Laplace is the Laplace distribution function, Δ is the feature sensitivity, and ∈ is the noise threshold.

[0151] Specifically, the local model data is generated by adding Laplace noise to the cropped sensor data using the noise addition formula. Schematically, the local model data H n It is to add noise to the cropped sensor data, add noise but keep the sensor data. is the cropped sensor data output after being clipped by the truncation function. Illustratively, the Laplace distribution function Laplace is used to generate noise that satisfies the input noise threshold. Optionally, the feature sensitivity Δ is the maximum difference between adjacent datasets of the cropped sensor data. Illustratively, the noise threshold ∈ is used to control noise intensity and can be set based on the user's noise privacy requirements.

[0152] This embodiment adds Laplace noise to the cropped sensor data after gradient clipping to obtain local model data. While retaining the data distribution trend, it is impossible to reversely infer the original data of multi-source sensors, thus achieving a balance between privacy and performance.

[0153] In one embodiment, based on a federated learning method and according to a secure aggregation protocol, local model data is globally fused to obtain shared warning data, including:

[0154] S701, obtaining global model data based on the local model data and the federated learning method;

[0155] Among them, the global model data is obtained using the following formula:

[0156]

[0157] Among them, G g is the global model data, N is the number of mines participating in federated learning, and H i is the i-th local model data.

[0158] Specifically, the local model data of multiple participating federated learning mines are globally integrated through the formula to obtain the global model data. Schematically, the global model data G g is the data obtained by fusing global and local model data. Optionally, the number of mines N participating in federated learning can be set by the user according to the actual situation of the mines. i It is the local model data output by the i-th mine after gradient clipping and noise addition processing.

[0159] S702, according to the security aggregation protocol, the global model data is globally integrated to obtain shared warning data;

[0160] The shared warning data is obtained using the following formula:

[0161] E(G g )=Ec(R,G g )

[0162] Among them, E(G g ) is the shared warning data, Ec is the encryption function, R is the public key, G g It is the global model data.

[0163] Specifically, based on the secure aggregation protocol, the global model data is encrypted using an encryption function and a set public key to generate shared warning data. g ) is the encrypted global model data, which cannot be directly parsed into the original model parameters, but supports ciphertext domain operations. Optionally, the encryption function Ec is an asymmetric encryption of the input data, which can be a Paillier homomorphic encryption algorithm. Schematically, the public key R is used for encryption operations and can be generated based on RSA or elliptic curve encryption algorithms. Optionally, the global model data G g It is the data obtained by fusing global and local model data.

[0164] This embodiment aggregates the local model data of each mine through a federated learning framework to obtain global model data, and adopts a secure aggregation protocol to perform homomorphic encryption on the global model data through public key encryption to obtain shared early warning data, thereby ensuring the security of shared early warning data during sharing, transmission and use, avoiding the leakage of commercial information, and providing training data for multi-platform early warning models, achieving the unity of privacy, security and model effectiveness.

[0165] like Figure 2 As shown, the embodiment of the present application provides a mine disaster monitoring data analysis and early warning method, which can also be called a mine disaster risk intelligent management and control method.

[0166] The following processes may be included:

[0167] Step S201: Multimodal data fusion and dynamic modeling.

[0168] Specifically, it includes: deploying a multi-source sensor network underground (vibration, gas, temperature and humidity, acoustic emission, etc.), coordinating monitoring of multiple types of sensors such as vibration, gas, temperature and humidity, and covering multi-dimensional disaster indicators of the mine environment; combining drone oblique photography and lidar scanning to build a millimeter-level three-dimensional geological model, and injecting key geological features such as faults, rock hardness, and groundwater distribution.

[0169] An improved Kalman filter algorithm is used to fuse sensor data. It eliminates noise caused by mine dust, electromagnetic interference and other complex noise scenarios, dynamically adapts to non-steady-state interference, and improves data accuracy.

[0170] Step S202: Intelligent disaster analysis and graded warning.

[0171] Specifically, it includes: realizing multi-dimensional data feature extraction and time series prediction based on the PSO-CNN-LSTM hybrid model; dynamically adjusting the convolution kernel parameters (such as quantity and size) of CNN and the number of hidden layer nodes of LSTM through particle swarm optimization (PSO), adaptively extracting the time series characteristics of sensor data and the spatial characteristics of geological models, and dynamically dividing disaster levels (low / medium / high / emergency).

[0172] By combining finite element analysis (FEA), geomechanical parameters (such as rock elastic modulus and permeability) with disaster characteristics (such as water inrush pressure and roof displacement), the propagation paths of disasters such as roof collapse and water inrush are simulated, and the impact range is predicted.

[0173] Step S203: Closed-loop response and resource coordination.

[0174] Specifically, it includes: hierarchical triggering of emergency mechanism: low-risk is pushed to the management terminal, and high-risk links the ventilation system, drainage equipment and evacuation navigation of personnel positioning terminals.

[0175] Based on blockchain technology, data sharing among multiple departments (such as rescue teams, hospitals, material warehouses, etc.) can be realized. Combined with the results of disaster simulation and prediction, cross-regional resource allocation paths can be dynamically optimized to shorten rescue response time and maximize rescue efficiency.

[0176] Step S204: self-optimization and knowledge iteration.

[0177] Specifically, this includes introducing a federated learning framework and, through a distributed parameter sharing mechanism, supporting the sharing of model parameters across mine nodes and global policy updates to adapt to dynamic mining environments while protecting the privacy of each mine's data. The real-time warning results of federated learning are fed back into the knowledge graph, dynamically updating the graph's node and edge weights.

[0178] Construct a disaster knowledge graph based on the historical disaster case database, associate multi-dimensional data such as geological characteristics, equipment status, and human operations, and realize intelligent tracing of the root causes of hidden dangers.

[0179] In the above-mentioned mine disaster monitoring data analysis and early warning method, by obtaining multi-source sensor data and the mine three-dimensional geological model in the mine, the multi-source sensor data and the mine three-dimensional geological model are fused based on the PSO-CNN-LSTM hybrid algorithm to obtain fused feature data and predicted disaster splitting levels. According to the classification level, the fused feature data is protected for security by using feature masking, gradient clipping and noise addition. Based on federated learning and security aggregation protocol, shared early warning data is generated to realize the sharing of mine sensor data for training comprehensive early warning models, thereby improving the accuracy of model disaster prediction.

[0180] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0181] Based on the same inventive concept, the embodiments of the present application also provide a mine disaster monitoring data analysis and early warning device for implementing the above-mentioned mine disaster monitoring data analysis and early warning method. The implementation solution provided by this device is similar to the implementation solution described in the above-mentioned method. Therefore, the specific limitations in one or more embodiments of the mine disaster monitoring data analysis and early warning device provided below can be referred to the limitations of the mine disaster monitoring data analysis and early warning method above, and will not be repeated here.

[0182] In an exemplary embodiment, Figure 3 As shown, a mine disaster monitoring data analysis and early warning device 300 is provided, comprising:

[0183] Information acquisition module 301, used to acquire multi-source sensor data and three-dimensional geological models of the mine;

[0184] The level determination module 302 is used to perform time series prediction and feature extraction on multi-source sensor data and three-dimensional geological models based on the PSO-CNN-LSTM hybrid algorithm to obtain disaster classification levels and fused feature data;

[0185] The information processing module 303 is used to perform feature masking on the fused feature data to obtain local model data when the disaster classification level meets the preset requirements;

[0186] The information sharing module 304 is used to globally fuse local model data based on the federated learning method and the security aggregation protocol to obtain shared warning data.

[0187] Furthermore, the level determination module includes:

[0188] Remove the high-frequency unit and use the following formula to calculate the high-frequency sensor data:

[0189]

[0190] in, is the high-frequency sensor data, DWT is the discrete wavelet transform function, z k It is to observe multi-source sensor data;

[0191] A covariance calculation unit, used to calculate the prediction covariance based on the high-frequency sensor data;

[0192] A noise removal unit is used to obtain denoised sensor data based on the removed high-frequency sensor data and the predicted covariance;

[0193] The denoised sensor data is calculated using the following formula:

[0194]

[0195] in, is the denoised sensor data at the current moment, is the denoised sensor data at the previous moment, K k is the Kalman gain, H k is the observation matrix, It is to remove high-frequency sensor data;

[0196] The model cutting unit is used to segment the three-dimensional geological model to obtain a set of geological model regions;

[0197] The feature fusion analysis unit is used to fuse and analyze the temporal characteristics of denoised sensor data and the spatial characteristics of the geological model area set based on the PSO-CNN-LSTM hybrid algorithm to obtain the disaster classification level and fused feature data.

[0198] Furthermore, the feature fusion analysis unit is also used to:

[0199] Obtaining denoised sensor data and a set of geological model regions;

[0200] Based on the LSTM algorithm, the candidate information state is obtained by constructing the input gate;

[0201] The candidate information state is obtained by the following formula:

[0202] C t =tanh(W C ·[h t-1 ,x t ]+b C )

[0203] Among them, C t is the candidate information state, W C is the input gate weight matrix, h t-1 is the hidden data of the previous moment, x t is the current input denoised sensor data, b c is the input bias term;

[0204] According to the candidate information state, the cell state information is updated to obtain the sensing time series data;

[0205] Among them, the sensing time series data is obtained through the following formula:

[0206] h t =o t ⊙tanh(C t )

[0207] Among them, h t is the sensor time series data, o t is the output hidden state data, ⊙ is the element-wise multiplication, C t It is to update information data;

[0208] Among them, the output hidden state data is obtained by the following formula:

[0209]

[0210] Among them, t is the output hidden state data, w0 is the output gate weight matrix, σ is the activation function, h t-1 is the hidden data of the previous moment, x t is the current input denoised sensor data, b0 is the output gate bias term;

[0211] Based on the CNN algorithm, the spatial characteristics of the geological model area set are extracted to obtain the geological characteristic map. The sensor time series data and the geological characteristic map are then comprehensively analyzed to obtain the disaster classification level and fusion feature data.

[0212] The calculation formula for disaster classification level is:

[0213] p=Softmax(W cls f fusion +b cls )

[0214] Where p is the disaster classification level, Softmax is the classification function, W cls is the classification weight, f fusion is the fusion feature data, b cls is the bias term;

[0215] Based on the PSO algorithm, the cross entropy loss is minimized and the convolution kernel parameters and LSTM hidden data dimension parameters are optimized.

[0216] Furthermore, the information processing module includes:

[0217] The feature masking unit is used to perform feature masking on the fused sensor data using the following formula to obtain masked sensor data:

[0218] X clean =X global ⊙M mask

[0219] Among them, X clean is to mask the sensor data, X global is the fused sensor data, ⊙ is the bitwise AND operator, M mask is the feature mask matrix;

[0220] A gradient calculation unit, configured to obtain gradient sensing data based on the mask sensing data and a time series analysis algorithm;

[0221] A gradient clipping unit, used for performing gradient clipping on the gradient sensing data to obtain clipped sensing data;

[0222] The noise adding unit is used to add noise to the cropped sensor data to obtain local model data.

[0223] Furthermore, the gradient clipping unit is also used to:

[0224] Use the following formula to perform gradient clipping on the gradient sensor data to obtain clipped sensor data:

[0225]

[0226] in, It is the cropped sensor data, Top is the interception function, is the gradient sensing data, and S is the interception ratio.

[0227] Furthermore, the noise adding unit is further configured to:

[0228] Use the following formula to add noise and obtain local model data:

[0229]

[0230] Among them, H n is the local model data, is the cropped sensor data, Laplace is the Laplace distribution function, Δ is the feature sensitivity, and ∈ is the noise threshold.

[0231] Furthermore, the information sharing module is also used to:

[0232] According to the local model data, the global model data is obtained based on the federated learning method;

[0233] Among them, the global model data is obtained using the following formula:

[0234]

[0235] Among them, G g is the global model data, N is the number of mines participating in federated learning, and H i is the i-th local model data;

[0236] According to the security aggregation protocol, the global model data is globally integrated to obtain shared warning data;

[0237] The shared warning data is obtained using the following formula:

[0238] E(G g )=Ec(R,G g )

[0239] Among them, E(G g ) is the shared warning data, Ec is the encryption function, R is the public key, G g It is the global model data.

[0240] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the mine disaster monitoring data analysis and early warning method as described above are implemented.

[0241] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0242] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the components described as separate parts may or may not be physically separated, and the parts displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the disclosed solution. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0243] The above-described embodiments merely represent several implementation methods of the embodiments of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the concept of the embodiments of the present application, and these modifications and improvements fall within the scope of protection of the embodiments of the present application.

Claims

1. A mine disaster monitoring data analysis and early warning method, characterized in that: The method comprises: Acquire multi-source sensor data and 3D geological models of mines; Based on the PSO-CNN-LSTM hybrid algorithm, time series prediction and feature extraction are performed on the multi-source sensor data and the three-dimensional geological model to obtain disaster classification levels and fusion feature data; When the disaster classification level meets the preset requirements, the fused feature data is subjected to feature masking to obtain local model data; Based on the federated learning method and in accordance with the security aggregation protocol, the local model data are globally fused to obtain shared warning data.

2. The method according to claim 1, characterized in that The PSO-CNN-LSTM hybrid algorithm is used to perform time series prediction and feature extraction on the multi-source sensor data and the three-dimensional geological model to obtain disaster classification levels and fusion feature data. include: The following formula is used to calculate the high-frequency sensor data: in, is the high-frequency sensor data, DWT is the discrete wavelet transform function, z k It is to observe multi-source sensor data; Calculating a prediction covariance based on the high-frequency-removed sensor data; Obtaining denoised sensor data according to the high-frequency-removed sensor data and the predicted covariance; The denoised sensor data is calculated using the following formula: in, is the denoised sensor data at the current moment, is the denoised sensor data at the previous moment, K k is the Kalman gain, H k is the observation matrix, It is to remove high-frequency sensor data; Segmenting the three-dimensional geological model to obtain a geological model region set; Based on the PSO-CNN-LSTM hybrid algorithm, the temporal characteristics of the denoised sensor data and the spatial characteristics of the geological model region set are fused and analyzed to obtain the disaster classification level and the fused feature data.

3. The method according to claim 2, characterized in that The PSO-CNN-LSTM hybrid algorithm is based on fusing and analyzing the temporal characteristics of the denoised sensor data and the spatial characteristics of the geological model region set to obtain the disaster classification level and the fused feature data, including: Acquiring the denoised sensor data and the geological model region set; Based on the LSTM algorithm, the candidate information state is obtained by constructing the input gate; The candidate information state is obtained by the following formula: C t =tanh(W C ·[h t-1 ,x t ]+b C ) Among them, C t is the candidate information state, W C is the input gate weight matrix, h t-1 is the hidden data of the previous moment, x t is the current input denoised sensor data, b c is the input bias term; According to the candidate information state, updating the cell state information to obtain sensing time series data; The sensing time series data is obtained by the following formula: h t =o t ⊙tanh(C t ) Among them, h t is the sensor time series data, o t is the output hidden state data, ⊙ is the element-wise multiplication, C t It is to update information data; Among them, the output hidden state data is obtained by the following formula: Among them, t is the output hidden state data, w0 is the output gate weight matrix, σ is the activation function, h t-1 is the hidden data of the previous moment, x t is the current input denoised sensor data, b0 is the output gate bias term; Based on the CNN algorithm, the spatial features of the geological model region set are extracted to obtain a geological feature map, and the sensor time series data and the geological feature map are comprehensively analyzed to obtain the disaster classification level and the fusion feature data; The calculation formula for the disaster classification level is: p=Softmax(W cls f fusion +b cls ) Where p is the disaster classification level, Softmax is the classification function, W cls is the classification weight, f fusion is the fusion feature data, b cls is the bias term; Based on the PSO algorithm, the cross entropy loss is minimized and the convolution kernel parameters and LSTM hidden data dimension parameters are optimized.

4. The method according to claim 1, wherein When the disaster classification level meets the preset requirements, the fused feature data is subjected to feature masking to obtain local model data, including: The fused sensor data is subjected to feature masking using the following formula to obtain masked sensor data: X clean =X global ⊙M mask Among them, X clean is to mask the sensor data, X global is the fused sensor data, ⊙ is the bitwise AND operator, M mask is the feature mask matrix; Obtaining gradient sensing data based on the masked sensing data and a time series analysis algorithm; Performing gradient clipping on the gradient sensing data to obtain the clipped sensing data; Noise is added to the cropped sensor data to obtain the local model data.

5. The method according to claim 4, characterized in that The step of performing gradient clipping on the gradient sensing data to obtain the clipped sensing data includes: The gradient sensing data is gradient clipped using the following formula to obtain the clipped sensing data: in, It is the cropped sensor data, Top is the interception function, is the gradient sensing data, and S is the interception ratio.

6. The method according to claim 4, characterized in that Adding noise to the cropped sensor data to obtain the local model data includes: Noise is added using the following formula to obtain the local model data: Among them, H n is the local model data, is the cropped sensor data, Laplace is the Laplace distribution function, Δ is the feature sensitivity, and ∈ is the noise threshold.

7. The method according to claim 1, characterized in that The federated learning method is based on a secure aggregation protocol, and the local model data is globally integrated to obtain shared warning data, including: Obtaining global model data based on the local model data and a federated learning method; The global model data is obtained using the following formula: Among them, G g is the global model data, N is the number of mines participating in federated learning, and H i is the i-th local model data; According to the security aggregation protocol, the global model data is globally integrated to obtain shared warning data; The shared warning data is obtained using the following formula: E(G g )=Ec(R,G g ) Among them, E(G g ) is the shared warning data, Ec is the encryption function, R is the public key, G g It is the global model data.

8. A mine disaster monitoring data analysis and early warning device, characterized in that: The device comprises: Information acquisition module, used to obtain multi-source sensor data and three-dimensional geological models of the mine; A level determination module is used to perform time series prediction and feature extraction on the multi-source sensor data and the three-dimensional geological model based on a PSO-CNN-LSTM hybrid algorithm to obtain a disaster classification level and fusion feature data; An information processing module, configured to perform feature masking on the fused feature data to obtain local model data when the disaster classification level meets preset requirements; The information sharing module is used to globally fuse the local model data based on the federated learning method and the security aggregation protocol to obtain shared warning data.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images