Method for predicting stability of large-span roof of mine cave experience pavilion

Abstract

The invention relates to artificial intelligence and data processing technologies, and discloses a mine cave experience pavilion large-span roof stability prediction method, which comprises the following steps: collecting data through sensor data collection, visual information collection and real-time environment data collection, and carrying out data fusion and cleaning; inputting sensor data, visual data and environmental disturbance data and outputting an instability risk level based on a multi-modal data prediction model; constructing a risk grade division standard according to the predicted instability risk grade; and dynamically updating the instability probability by using a Bayesian method in combination with historical data and real-time data. The large-span roof safety risk prediction method can accurately predict the safety risk of a large-span roof, so that potential safety risks and hidden dangers can be found in time, hidden dangers can be eliminated in time or personnel and property are transferred, and personnel and property safety is guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method for predicting the stability of a large-span roof of a mine experience hall. Background Art

[0002] The mine experience center project was rebuilt on the basis of an abandoned mine. In order to improve the tourist experience, two large-span curved roof venue buildings were built outside the mine. Since the stability of the large-span roof structure is crucial, in order to ensure the safety and stability of the structure, a large-span roof stability prediction method for the mine experience center was designed to detect and predict the stability of the large-span roof, and to promptly discover potential safety risks and hidden dangers, so as to promptly eliminate hidden dangers or transfer personnel and property. Summary of the invention

[0003] The present invention provides a method for predicting the stability of a large-span roof of a mine experience hall, the main purpose of which is to more accurately predict the safety risks of the large-span roof, so as to timely discover potential safety risks and hidden dangers, so as to timely eliminate hidden dangers or transfer personnel and property to ensure the safety of personnel and property.

[0004] To achieve the above object, the present invention provides a method for predicting the stability of a large-span roof of a mine experience hall, comprising: S1 Data Collection and Processing Sensor network layout: Multiple types of sensors are deployed at key structural nodes of large-span roofs, including truss nodes and stress-concentrated locations. Multiple types of sensors include strain gauges, accelerometers, and temperature and humidity sensors. Use drones and ground-based LiDAR to collect visual information on roof deformation and cracks; Real-time environmental data collection: Acquisition of meteorological data, including wind speed, wind direction, and temperature changes, as external disturbance factors affecting roof stability; Data fusion and cleaning: Use signal processing technology to synchronize multi-source data and eliminate outliers; S2 Prediction model based on multimodal data Model structure design: Step 2.1 Input layer: Sensor data: real-time monitoring data of stress, strain, and displacement of key nodes on the roof; visual data: cracks on the surface of the structure, local deformation characteristics; environmental disturbance data: meteorological data including wind load, snowfall, temperature and humidity; Step 2.2 Feature extraction layer: extract the deformation features of the image through convolutional neural network; use long short-term memory network to model the dynamic features of sensor time series data; Step 2.3 Fusion layer: Use multimodal deep learning models to fuse various data features; Step 2.4 Output layer: Output the instability risk level, which is divided into low, medium and high; S3) Early warning mechanism According to the predicted instability risk level, a risk level classification standard is constructed: normal-safe, warning-needs attention, critical-needs intervention; combining historical data with real-time data, the Bayesian method is used to dynamically update the instability probability.

[0005] In the above method, data fusion and cleaning include: Data preprocessing: synchronize the time and frequency of the collected multi-source data; interpolate missing values ​​to ensure data continuity; Outlier detection: Apply statistical characteristics and signal characteristics methods to conduct preliminary outlier screening; use machine learning methods to further analyze the screened abnormal data to ensure the accuracy of elimination; Outlier elimination: Eliminate data that is determined to be outliers; re-interpolate or smooth the eliminated data to ensure signal integrity; Result verification: Compare the quality of the signal before and after elimination to verify the processing effect; verify the impact of data processing on prediction accuracy in practical applications.

[0006] In the above method, convolutional neural network: extracts image deformation features 1) Input data Image data: images of large-span roof structures generated by drones, surveillance cameras or LiDAR; Image content: Crack changes, including width and length; deformation characteristics, including roof bulge and subsidence; signs of structural damage, including rust and deformed nodes; 2) Network structure of convolutional neural network Convolutional neural networks are used to extract local features and global structural features in images. The specific network design is as follows: Convolutional layer: Use multiple convolution kernels to extract edge, texture, and local deformation features; the convolution kernel of each layer gradually expands the receptive field and extracts higher-level shape and deformation information; Formula: The output feature map size is ; in, H in ,、 W in is the input image size, K is the convolution kernel size, P is filling, S is the step length; Pooling layer: Use maximum pooling or average pooling to reduce the dimension of features, reduce the amount of calculation and improve the robustness of the model; Fully connected layer: maps high-dimensional feature vectors into compact low-dimensional representations; extracts global features of the image, such as deformation feature distribution; 3) Feature Type Low-level features: edge, crack texture; High-level features: overall deformation mode of large-span roofs, including local depression or overall tilt; 4) Output Image feature vector, used as input to the fusion layer.

[0007] In the above methods, Transformer or self-attention mechanism is used to realize multimodal feature fusion. The workflow of the fusion layer is: 1) Input Data features from different modalities: Image features: feature vectors extracted by convolutional neural networks; Time series features: feature vectors extracted by long short-term memory networks; These features are uniformly represented as a multimodal feature matrix: F=[F image ,F time ,F other ] Among them, F image ,F time ,F other are the eigenvectors of different modes respectively; 2) Feature alignment If the feature vectors have different dimensions, first align them to the same dimension through a fully connected layer or linear transformation d ; 3) Fusion Mechanism Use Transformer or self-attention mechanism to fuse the aligned features and model the global dependencies between modalities; 4) Output A fused global feature vector is used as input for downstream tasks.

[0008] In the above methods, feature fusion based on self-attention mechanism: 1) Self-attention mechanism formula The calculation of self-attention is divided into the following steps: Input feature representation: Given a multimodal feature matrix F ∈ R n×d ,in n is the number of modes, d is the characteristic dimension; Linear mapping: Mapping input features to query, key, and value vectors: Q = FW Q , K = FW K, V = FW V Among them, W Q ,W K ,W V ∈R d×dk is the learnable weight matrix, d k is the dimension of the low-dimensional feature space; Attention score calculation: Calculate the similarity between the query and the key through the dot product, and get the attention weight through Softmax: ; Among them, A∈R n×n Represents the correlation between modes.

[0009] Weighted summation: Use the attention weight A and the weighted value vector V to get the fused feature representation: F'=AV 2) Multi-head self-attention mechanism Use multiple attention heads to calculate in parallel to capture the feature relationships of different subspaces: MultiHead(Q,K,V)=Concat(head1,…,head h )W O Among them, head i = Attention(Q i , K i , V i ), W O is the output weight matrix.

[0010] In the above methods, Transformer-based multimodal fusion: 1) Transformer basic structure Transformer consists of the following main components: a multi-head self-attention layer to capture the global interaction between modalities; a feedforward network to perform nonlinear transformations on the output of the attention layer to enhance feature representation capabilities; residual connections and normalization to keep gradients stable and avoid the vanishing gradient problem; 2) Multimodal fusion process Input representation: multimodal features F =[ F image , F time ] are concatenated into a matrix with the shape n×d ; Multi-head self-attention layer: calculates the attention relationship between modalities and obtains the fused feature representation; Feedforward network: further feature transformation of attention output through two layers of fully connected networks; Output features: Finally output a global feature vector F fusion , used for downstream tasks.

[0011] The main advantages of the present invention are as follows: 1) Efficiency of dynamic stability prediction The Bayesian method combines historical data with real-time sensor data to dynamically update the probability of instability; evaluate the impact of environmental factors (such as wind load, snow load, temperature and humidity) in real time, and quickly adjust the prediction results; iteratively improve, and as time goes by, the increase in observation data continuously optimizes prior knowledge and improves prediction accuracy.

[0012] 2) Comprehensiveness of multimodal data fusion Comprehensive utilization of information is achieved by integrating visual data (such as crack images), time series data (sensor data such as strain and vibration), and environmental data (wind, temperature and humidity, etc.); with the help of deep learning (such as CNN to extract image features and LSTM to model time series data) and self-attention mechanism, the correlation between multimodal data is extracted to improve the performance of the prediction model.

[0013] 3) Strong robustness and adaptability Combining visual monitoring and physical sensors can avoid the impact of failure of a single data source on prediction results; the self-attention mechanism can dynamically adjust weights under data noise or abnormal conditions to improve the robustness of the model; and it can adapt to dynamic changes in different environments.

[0014] The large-span roof stability prediction method for a mine cave experience hall proposed in the present invention can more accurately predict the stability of a large-span roof structure, output the instability risk level through the model, provide intuitive risk assessment information for maintenance and management personnel, provide accurate risk prediction, and reduce unnecessary maintenance costs and waste of human resources; reduce the safety hazards caused by structural instability, ensure the long-term operation of the mine cave experience hall, and promptly discover potential safety risks and hidden dangers so as to promptly eliminate the hidden dangers or transfer personnel and property to ensure the safety of personnel and property. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 The figure is a schematic diagram of the flow chart of the prediction method of the present invention. DETAILED DESCRIPTION

[0016] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings.

[0017] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0018] The embodiment of the present application provides a method for predicting the stability of a large-span roof of a mine cave experience hall. The execution subject of the prediction method includes but is not limited to at least one of the electronic devices such as a server and a terminal that can be configured to execute the method provided by the embodiment of the present application. In other words, the method for predicting the stability of a large-span roof of a mine cave experience hall can be executed by software or hardware installed on a terminal device or a server device, and the software can be a blockchain platform. The server includes but is not limited to: a single server, a server cluster, a cloud server or a cloud server cluster, etc. The server can be an independent server, or it can be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (Content Delivery Network, CDN), and big data and artificial intelligence platforms.

[0019] Reference Figure 1 FIG. 1 is a flow chart of a method for predicting the stability of a large-span roof of a mine cave experience hall provided by an embodiment of the present invention. In this embodiment, the method for predicting the stability of a large-span roof of a mine cave experience hall includes: S1. Data collection and processing Sensor network layout: Multiple types of sensors are deployed at key structural nodes of large-span roofs. The key structural nodes include at least truss nodes and stress-concentrated locations. The multiple types of sensors include at least strain gauges, accelerometers, and temperature and humidity sensors. Use drones and ground-based LiDAR to collect visual information on roof deformation and cracks; Real-time environmental data collection: Acquisition of meteorological data, including wind speed, wind direction, and temperature changes, as external disturbance factors affecting roof stability; Data fusion and cleaning: Use signal processing technology to synchronize multi-source data and eliminate outliers; Data fusion and cleaning include: Data preprocessing: synchronize the time and frequency of the collected multi-source data; interpolate missing values ​​to ensure data continuity; Outlier detection: Apply statistical characteristics and signal characteristics methods to conduct preliminary outlier screening; use machine learning methods to further analyze the screened abnormal data to ensure the accuracy of elimination; Outlier elimination: Eliminate data that is determined to be outliers; re-interpolate or smooth the eliminated data to ensure signal integrity; Result verification: Compare the quality of the signal before and after elimination to verify the processing effect; verify the impact of data processing on prediction accuracy in practical applications.

[0020] Signal processing technology for synchronizing multi-source data and eliminating outliers needs to be combined with data acquisition characteristics and application scenarios: 1) Synchronous processing technology Multi-source data may come from different types of sensors (such as strain, vibration, temperature and humidity sensors) or devices with different sampling frequencies, and need to be synchronized through time alignment, frequency alignment and data interpolation.

[0021] Step 1.1 Time alignment Use timestamp unification: Ensure that all sensor data is time-stamped with high precision (such as NTP time synchronization).

[0022] If there is a timestamp deviation, a time difference correction method (such as a maximum cross-correlation function method) is used to align the timestamps of different sensor data.

[0023] Steps: Collect the original timestamp of each sensor data; align the time of the reference sensor and adjust the data of other sensors to the same time axis.

[0024] Step 1.2 Frequency Alignment Upsampling / Downsampling: Use signal resampling methods to unify the data to the target sampling frequency.

[0025] Commonly used methods: linear interpolation, Lagrange interpolation or spline interpolation (suitable for nonlinear changes).

[0026] Bandpass filtering: Filter high-frequency or low-frequency signals to retain valid information within the target frequency band and eliminate redundant frequency components.

[0027] Step 1.3 Data interpolation Function: Supplement missing values ​​or homogenize non-uniformly sampled data.

[0028] Linear interpolation: suitable for situations where data changes slowly.

[0029] Spline interpolation: suitable for nonlinear dynamic signals.

[0030] Kalman filter interpolation: Smoothing noisy data while interpolating missing values.

[0031] 2) Outlier removal technology Outliers may be caused by sensor failure, external interference or extreme environmental conditions. According to the characteristics of different outliers, the following elimination methods are used: Step 2.1 Outlier detection based on statistical characteristics method: Z-score method: The data were normalized and the Z score (i.e. the standard deviation distance from the mean) was calculated for each data point.

[0032] Formula: Z = (x − μ ) / σ If |Z|> threshold} (such as 2 or 3), it is determined to be an outlier.

[0033] Box plot method: The outlier range was defined based on the interquartile range (IQR).

[0034] Outlier judgment formula: lower than Q1 -1.5× IQR or higher Q3 +1.5× IQR The points are considered as outliers.

[0035] Step 2.2 Outlier detection based on signal characteristics Threshold method: According to the sensor working range or theoretical model, the physical upper and lower limits of the data are set, and the values ​​outside the range are considered abnormal; Sliding window mean detection: Calculate the sliding mean of the signal and determine whether the deviation between the current data point and the sliding mean exceeds the preset threshold; Fourier transform: Perform frequency domain analysis on time domain signals to remove abnormal frequency components in the spectrum that do not meet expectations; Wavelet transform: Use wavelet transform to analyze the time-frequency characteristics of the signal and identify mutation points or abnormal frequency bands; Step 2.3 Outlier Removal Based on Machine Learning Isolation Forest: It randomly splits the data to determine whether the data point is isolated. Principal Component Analysis (PCA): After dimensionality reduction, the projection of the signal in the principal component space is analyzed. Signals whose projection values ​​are far away from other points are considered outliers. Deep learning: train autoencoders to detect outliers based on reconstruction errors; apply recurrent neural networks (RNNs) or long short-term memory networks (LSTMs) to model time series data and identify abnormal behaviors; Step 2.4 Real-time culling technology Sliding window real-time processing: Use sliding windows on data streams to detect outliers in real time to avoid delays; Robust filtering: Use median filtering or Kalman filtering to remove outliers in real time.

[0036] Through synchronization processing and outlier elimination technology, a reliable and clean data foundation is provided for the subsequent large-span roof stability prediction model, thereby improving the accuracy and robustness of the model.

[0037] S2 Prediction model based on multimodal data Model structure design: 1) Input layer: sensor data: real-time monitoring data of key roof nodes including stress, strain, and displacement; visual data: structural surface cracks and local deformation characteristics; environmental disturbance data: meteorological data including wind load, snowfall, temperature and humidity; 2) Feature extraction layer: extract the deformation features of the image through convolutional neural network (CNN); Step 2.1 Convolutional Neural Network (CNN): Extracting Image Deformation Features Step 2.1.1 Input data Image data: images of large-span roof structures generated by drones, surveillance cameras or LiDAR; Image content: Crack changes, including width and length; deformation characteristics, including roof bulge and subsidence; signs of structural damage, including rust and deformed nodes; Step 2.1.2 CNN network structure CNN is used to extract local features and global structural features in images. The specific network design is as follows: Convolutional layer: Use multiple convolution kernels to extract edge, texture, and local deformation features; the convolution kernel of each layer gradually expands the receptive field and extracts higher-level shape and deformation information; Formula: The output feature map size is ; in, H in ,、 W in is the input image size, K is the convolution kernel size, P is filling, S is the step length; Pooling layer: Use maximum pooling or average pooling to reduce the dimension of features, reduce the amount of calculation and improve the robustness of the model; Fully connected layer: maps high-dimensional feature vectors into compact low-dimensional representations; extracts global features of the image, such as deformation feature distribution; Step 2.1.3 Feature Type Low-level features: edge, crack texture; High-level features: overall deformation mode of large-span roofs, including local depression or overall tilt; Step 2.1.4 Output Image feature vector, used as input to the fusion layer.

[0038] Step 2.2 Use the long short-term memory network (LSTM) to model the dynamic characteristics of sensor time series data; The long short-term memory network (LSTM) is used to model the dynamic characteristics of sensor time series data, mainly to capture the dynamic change characteristics of long-term dependence in the time series (such as trends, periodic changes, mutations, etc.). The following are the complete design, implementation and process details: Step 2.2.1 LSTM modeling: Analyze the time series data collected by sensors (such as stress, strain, vibration, temperature and humidity changes, etc.).

[0039] Capture dynamic features, including: short-term features: such as rapidly changing vibration signals and transient responses; long-term features: such as gradually accumulated strain or trend changes under long-term loads.

[0040] Predict specific outputs: Future states (e.g. deformation amplitude, vibration signal strength). Anomaly detection (e.g. large changes that may cause structural instability).

[0041] Sensor data sources: Acceleration sensor: vibration signal; strain gauge: local strain time series; temperature and humidity sensor: impact of external environmental disturbance on the structure.

[0042] The typical LSTM network architecture includes: input layer: receiving sensor time series data; hidden layer: composed of LSTM units, used to extract dynamic features; output layer: mapping hidden states to predicted targets.

[0043] Step 2.2.2 LSTM modeling process: Data preparation Time series data format: The input data is a multidimensional time series matrix: ; T: time step; N: dimension of sensor data.

[0044] Standardization: Normalize or standardize the data (such as Min-Max normalization) to speed up model convergence.

[0045] Window segmentation: Split the dataset into time windows (such as sliding window method) for training and prediction: Input: [x t ,x t+1 ,…,x t+w ].

[0046] Output: x t+w+1 (next time step prediction).

[0047] Step 2.2.3 Network structure design Input Layer: The input shape is (batch_size, seq_length, input_dim).

[0048] seq_length: The window size in time steps.

[0049] input_dim: Feature dimension per time step (number of sensor types).

[0050] LSTM Layer: Hidden layer dimension (hidden_size): defines the dimension size of the hidden state (such as 128).

[0051] Number of layers (num_layers): stack multiple LSTM layers (such as 2 layers).

[0052] Fully connected layer: Map the output of LSTM to a specific prediction target (such as time series point value or classification).

[0053] Step 2.3 Combination of CNN and LSTM: Multimodal Feature Extraction In order to comprehensively process image data and time series data, CNN and LSTM are combined and feature fused in the following ways: The image features extracted by CNN (such as deformation feature vectors) are fused with the temporal features extracted by LSTM (such as dynamic change feature vectors).

[0054] Fusion method: Serial fusion: directly concatenate feature vectors to form a joint feature representation: F fusion =[F image ,F time ] Weighted fusion: assign weights to different features: F fusion =αF image +βF time F fusion is the fused feature, F image Image features extracted by CNN, F time is the time series feature extracted by LSTM, α、β is the feature weight; Attention mechanism: Dynamically adjust feature weights through the self-attention model.

[0055] Shared fusion layer: Build a shared fully connected layer to receive feature inputs from CNN and LSTM for joint learning and optimization. Through this design, complementary features can be extracted from visual data and time series data.

[0056] 3) Fusion layer: Utilize multimodal deep learning models to fuse various data features. The workflow of the fusion layer is as follows: Step 3.1 Input Data features from different modalities: Image features: feature vectors (e.g., 256 dimensions) extracted by convolutional neural networks (CNNs); Time series features: feature vectors (e.g., 128 dimensions) extracted by long short-term memory networks (LSTM); These features are uniformly represented as a multimodal feature matrix: F=[F image ,F time ,F other ] Among them, F image ,F time ,F other are the eigenvectors of different modes respectively; Step 3.2 Feature Alignment If the dimensions of the feature vectors are different, first align them to the same dimension through a fully connected layer (or linear transformation) d ; Step 3.3 Fusion mechanism Use Transformer or self-attention mechanism to fuse the aligned features and model the global dependencies between modalities; Step 3.4 Output A fused global feature vector is used as input for downstream tasks.

[0057] Step 3.5 Feature fusion based on self-attention mechanism: Step 3.5.1 Self-Attention Mechanism Formula The calculation of self-attention is divided into the following steps: Input feature representation: Given a multimodal feature matrix F ∈ R n×d ,in n is the number of modes, d is the characteristic dimension; Linear mapping: Mapping input features to query, key, and value vectors: Q = FW Q , K = FW K , V = FW V Among them, W Q ,W K ,W V ∈R d×dk is the learnable weight matrix, d k is the dimension of the low-dimensional feature space; Attention score calculation: Calculate the similarity between the query and the key through the dot product, and get the attention weight through Softmax: ; Among them, A∈R n×n Represents the correlation between modes.

[0058] Weighted summation: Use the attention weight A and the weighted value vector V to get the fused feature representation: F'=AV Step 3.5.2 Multi-head self-attention mechanism Use multiple attention heads (Multi-Head Attention) for parallel calculation to capture the feature relationships of different subspaces: MultiHead(Q,K,V)=Concat(head1,…,head h )W O Among them, head i = Attention(Q i , K i , V i ), W O is the output weight matrix.

[0059] Step 3.6 Transformer-based multimodal fusion: Step 3.6.1 Transformer basic structure Transformer consists of the following main components: a multi-head self-attention layer to capture the global interaction between modalities; a feedforward network to perform nonlinear transformations on the output of the attention layer to enhance feature representation capabilities; residual connections and normalization to keep gradients stable and avoid the vanishing gradient problem; Step 3.6.2 Multimodal fusion process Input representation: multimodal features F =[ F image , F time ] are concatenated into a matrix with the shape n×d ; Multi-head self-attention layer: calculates the attention relationship between modalities and obtains the fused feature representation; Feedforward network: further feature transformation of attention output through two layers of fully connected networks; Output features: Finally output a global feature vector F fusion , used for downstream tasks.

[0060] 4) Output layer: Based on the input data, predict whether the instability risk belongs to a certain level (low, medium, high).

[0061] Output form: classification probability, indicating the probability distribution of each risk level.

[0062] Risk level classification task method: The output features of the fusion layer are mapped to the dimension of the number of classification categories through a fully connected layer.

[0063] Activation function: Use Softmax to convert the output into a probability distribution.

[0064] formula: ; in: P (class i ): probability of the i-th category; z i : The predicted score of the ith category; C: The number of classification categories (3 categories in total: low, medium and high).

[0065] Python: class RiskClassificationOutput(nn.Module): def __init__(self, input_dim, num_classes): super(RiskClassificationOutput, self).__init__() self.fc = nn.Linear(input_dim, num_classes) self.softmax = nn.Softmax(dim=1) def forward(self, x): x = self.fc(x) # Output category score x = self.softmax(x) # Convert to probability distribution return x Input: fused feature vector (eg 128 dimensions).

[0066] Output: Class probability distribution (shape is ( batch_size, num_classes )).

[0067] S3) Early warning mechanism According to the predicted instability risk level, a risk classification standard is established: normal (safe), warning (needs attention), critical (needs intervention); historical data and real-time data are combined to dynamically update the instability probability using the Bayesian method; a multi-level warning mechanism is set up to send warning information in real time through the Internet of Things platform, for example, on-site alarms, sending mobile phone text messages, phone calls, WeChat and other warning methods.

[0068] By combining historical data (prior information) with real-time data (observation information), the probability of instability is dynamically updated: Step 3.1 Basic form of Bayesian formula The Bayesian formula is as follows: ; P(H|D): posterior probability, the probability of the unstable state H given the real-time data D (the target to be calculated).

[0069] P(H): Prior probability, the initial probability of the unstable state H estimated based on historical data and empirical knowledge.

[0070] P(D|H): Likelihood function, the probability of observing real-time data D under the unstable state H (determined by the real-time data and the model).

[0071] P(D): Marginal probability, the overall probability of observing data D, which plays a normalizing role.

[0072] Step 3.2 Model construction Step 3.2.1 Define the instability state H. Assume that the roof structure has two states: H=1: structural instability; H=0: structural stability.

[0073] Step 3.2.2 Prior probability P(H) Determined based on historical data (such as the probability of instability of similar structures under similar load and environmental conditions).

[0074] Example: If the probability of instability is estimated to be 10% based on historical data, then P(H=1)=0.1, P(H=0)=0.9 Step 3.2.3 Likelihood function P(D|H) Describes the distribution of real-time data D under a given state H.

[0075] Determined through physical models, sensor observations, or statistical learning models.

[0076] For example: Observation data D: stress, strain, crack width, vibration frequency, etc.

[0077] Assume that when H=1 (instability), the distribution of a certain eigenvalue x is Gaussian: ; When H=0 (stable), the distribution of this eigenvalue is: ; Step 3.2.4 Marginal Probability P(D) Calculated by the total probability formula: ; Step 3.3 Dynamic update of real-time instability probability Step 3.3.1 Update formula Observing data in real time D Based on the dynamic update of the instability probability P(H = 1|D) : ; Step 3.3.2 Iterative update of multiple data points If the data stream D = {D1,D2,…,D t}Continuous input, the posterior probability can be gradually updated using the Bayesian formula: Initial prior: P(H = 1) .

[0078] No. t The posterior probability at time: ; in , P(H = 1|D t-1 ) is the posterior probability of the previous moment; P(D t |H=1) is the likelihood of real-time data.

[0079] The dynamic Bayesian model can dynamically update the instability probability based on new observation data; based on existing historical data, the instability prediction can be corrected through real-time data; and probabilistic results can be provided to facilitate decision-making under uncertain conditions.

[0080] It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0081] Therefore, no matter from which point of view, the embodiments should be regarded as illustrative and non-restrictive, and the scope of the present invention is limited by the appended claims rather than the above description, so it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any attached figure mark in the claims should not be regarded as limiting the claims involved.

[0082] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Artificial Intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0083] In addition, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or systems stated in a system claim can also be implemented by one unit or system through software or hardware. The words first, second, etc. are used to indicate names, and do not indicate any particular order.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for predicting the stability of a large-span roof of a mine experience hall, characterized in that: The method comprises: S1 Data Collection and Processing Sensor network layout: Multiple types of sensors are deployed at key structural nodes of large-span roofs, including truss nodes and stress-concentrated locations. Multiple types of sensors include strain gauges, accelerometers, and temperature and humidity sensors. Use drones and ground-based LiDAR to collect visual information on roof deformation and cracks; Real-time environmental data collection: Acquisition of meteorological data, including wind speed, wind direction, and temperature changes, as external disturbance factors affecting roof stability; Data fusion and cleaning: Use signal processing technology to synchronize multi-source data and eliminate outliers; Use Transformer or self-attention mechanism to realize multimodal feature fusion. The workflow of the fusion layer is as follows: Step 1 Input Data features from different modalities: Image features: feature vectors extracted by convolutional neural networks; Time series features: feature vectors extracted by long short-term memory networks; These features are uniformly represented as a multimodal feature matrix: F=[F image ,F time ,F other ] Among them, F image ,F time ,F other are the eigenvectors of different modes respectively; Step 2: Feature alignment If the feature vectors have different dimensions, first align them to the same dimension through a fully connected layer or linear transformation. d ; Step 3: Fusion Mechanism Use Transformer or self-attention mechanism to fuse the aligned features and model the global dependencies between modalities; Feature fusion based on self-attention mechanism: 1) Self-attention mechanism formula The calculation of self-attention is divided into the following steps: Input feature representation: Given a multimodal feature matrix F∈R n×d ,in n is the number of modes, d is the characteristic dimension; Linear mapping: Mapping input features to query, key, and value vectors: Q = FW Q ,K=FW K ,V=FW V Among them, W Q ,W K ,W V ∈R d×dk is the learnable weight matrix, d k is the dimension of the low-dimensional feature space; Attention score calculation: Calculate the similarity between the query and the key through the dot product, and get the attention weight through Softmax: ； Among them, A∈R n×n Represents the correlation between modes; Weighted summation: Use the attention weight A and the weighted value vector V to get the fused feature representation: F'=AV 2) Multi-head self-attention mechanism Use multiple attention heads to calculate in parallel to capture the feature relationships of different subspaces: MultiHead(Q,K,V)=Concat(head1,…,head h )W O Among them, head i = Attention(Q i , K i , V i ), W O is the output weight matrix; Step 4 Output A fused global feature vector, used as input for downstream tasks; S2 Prediction model based on multimodal data Model structure design: Step 2.1 Input layer: Sensor data: real-time monitoring data of stress, strain, and displacement of key nodes on the roof; visual data: cracks on the surface of the structure, local deformation characteristics; environmental disturbance data: meteorological data including wind load, snowfall, temperature and humidity; Step 2.2 Feature extraction layer: extract the deformation features of the image through convolutional neural network; use long short-term memory network to model the dynamic features of sensor time series data; Step 2.3 Fusion layer: Use multimodal deep learning models to fuse various data features; Step 2.4 Output layer: Output the instability risk level, which is divided into low, medium and high; S3 early warning mechanism According to the predicted instability risk level, a risk level classification standard is constructed: normal-safe, warning-needs attention, critical-needs intervention; combining historical data with real-time data, the Bayesian method is used to dynamically update the instability probability.

2. A method for predicting stability of a large-span roof of a mine cave experience hall as claimed in claim 1, characterized in that: Data fusion and cleaning include: Data preprocessing: synchronize the time and frequency of the collected multi-source data; interpolate missing values ​​to ensure data continuity; Outlier detection: Apply statistical characteristics and signal characteristics methods to conduct preliminary outlier screening; use machine learning methods to further analyze the screened abnormal data to ensure the accuracy of elimination; Outlier elimination: Eliminate data that is determined to be outliers; re-interpolate or smooth the eliminated data to ensure signal integrity; Result verification: Compare the quality of the signal before and after elimination to verify the processing effect; verify the impact of data processing on prediction accuracy in practical applications.

3. A method for predicting stability of a large-span roof of a mine cave experience hall as claimed in claim 1, characterized in that: Convolutional Neural Network: Extracting Image Deformation Features 1) Input data Image data: images of large-span roof structures generated by drones, surveillance cameras or LiDAR; Image content: crack changes, including width and length; Deformation characteristics, including roof bulge and subsidence; Signs of structural damage, including corrosion and deformed joints; 2) Network structure of convolutional neural network Convolutional neural networks are used to extract local features and global structural features in images. The specific network design is as follows: Convolutional layer: Use multiple convolution kernels to extract edge, texture, and local deformation features; the convolution kernel of each layer gradually expands the receptive field and extracts higher-level shape and deformation information; Formula: The output feature map size is ； in, H in ,、 W in is the input image size, K is the convolution kernel size, P is filling, S is the step length; Pooling layer: Use maximum pooling or average pooling to reduce the dimension of features, reduce the amount of calculation and improve the robustness of the model; Fully connected layer: maps high-dimensional feature vectors into compact low-dimensional representations; extracts global features of the image, such as deformation feature distribution; 3) Feature Type Low-level features: edge, crack texture; High-level features: overall deformation mode of large-span roofs, including local depression or overall tilt; 4) Output Image feature vector, used as input to the fusion layer.

4. A method for predicting stability of a large-span roof of a mine cave experience hall as claimed in claim 1, characterized in that: Transformer-based multimodal fusion: 1) Transformer basic structure Transformer consists of the following main components: a multi-head self-attention layer to capture the global interaction between modalities; a feedforward network to perform nonlinear transformations on the output of the attention layer to enhance feature representation capabilities; residual connections and normalization to keep gradients stable and avoid the vanishing gradient problem; 2) Multimodal fusion process Input representation: multimodal features F =[ F image , F time ] are concatenated into a matrix with the shape n×d ; Multi-head self-attention layer: calculates the attention relationship between modalities and obtains the fused feature representation; Feedforward network: further feature transformation of attention output through two layers of fully connected networks; Output features: Finally output a global feature vector F fusion , used for downstream tasks.

Images