An adaptive migration learning disaster prediction method under complex geological conditions

Abstract

The application discloses a kind of self-adapting migration learning disaster prediction methods under complex geological conditions, it is related to intelligent disaster prediction technical field, including, the multi-source heterogeneous data of target mine is collected, and multi-source heterogeneous data is washed, denoised and standardized, utilizes multi-source heterogeneous data and trains initial prediction model, obtains disaster prediction model;Disaster prediction model is deployed to edge computing server, and data connection is established with real-time sensor network, executes forward calculation to new multi-source heterogeneous data, and outputs preliminary risk situation spectrum;Real-time data sequence fragments in preliminary risk situation spectrum are acquired, real-time data sequence fragments are input into lightweight time series Transform network, and the difference degree between time series mode and stable mode is calculated, when difference degree exceeds sensitivity threshold, migration learning trigger signal is generated;The application improves prediction accuracy and reliability, provides efficient, real-time support for disaster management.

Description

Technical Field

[0001] This invention relates to the field of intelligent disaster prediction technology, and in particular to an adaptive transfer learning disaster prediction method under complex geological conditions. Background Technology

[0002] Mine disaster prediction technology began with empirical threshold discrimination and statistical analysis of key monitoring quantities such as gas concentration, roof delamination, surrounding rock stress, and microseismic events. It has gradually developed into physical models based on geomechanical theory and rock mass failure mechanisms, and statistical prediction models represented by regression analysis and time series analysis. With the construction of the Internet of Things (IoT) and the Industrial Internet in mines, multi-source sensors such as gas sensors, stress-strain sensors, and seismic detectors have been widely deployed underground, forming a multi-source monitoring network covering mining faces and roadways. Based on this, existing technologies have proposed multi-source information fusion methods for mine disaster risk assessment. By cleaning, filtering, and extracting features from sensor data, machine learning or deep learning models such as support vector machines, random forests, and deep neural networks are used to establish probabilistic prediction models for disasters such as mine stress manifestation, rockburst, and gas outburst. These models are then deployed at ground control centers or edge computing nodes to achieve rolling assessments of mine disaster risks.

[0003] Under complex geological conditions and frequently changing mining conditions, the statistical distribution and temporal evolution characteristics of monitoring data often undergo "conceptual drift" with factors such as increased mining depth, changes in geological structure, and adjustments in development methods. This invention proposes an adaptive transfer learning disaster prediction method under complex geological conditions. It utilizes multi-source heterogeneous data to construct a disaster prediction model and introduces a lightweight temporal Transformer in an edge computing environment to perform pattern analysis on the preliminary risk situation spectrum. The method uses difference-driven transfer learning to trigger and construct a new conditional sample set from real-time data. It then performs small-step fine-tuning and mask-selective activation on the deep temporal prediction network, thereby providing a solution to the main problem of insufficient adaptability of the above-mentioned model and the secondary problem of the need to improve the precision of risk pattern recognition. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an adaptive transfer learning disaster prediction method under complex geological conditions to solve the problem that disaster prediction models are difficult to adapt to changes in complex geology and working conditions in a timely manner and are not sensitive enough to the identification of changes in risk situation patterns.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides an adaptive transfer learning disaster prediction method under complex geological conditions, which includes collecting multi-source heterogeneous data of the target mine, cleaning, denoising and standardizing the multi-source heterogeneous data, training an initial prediction model using the multi-source heterogeneous data, and obtaining a disaster prediction model.

[0008] The disaster prediction model is deployed to an edge computing server and a data connection is established with a real-time sensor network. Forward calculations are performed on new multi-source heterogeneous data to output a preliminary risk situation spectrum.

[0009] The system acquires real-time data sequence fragments from the preliminary risk situation spectrum, inputs these fragments into a lightweight temporal Transformer network, and calculates the difference between the temporal pattern and the stable pattern. When the difference exceeds the sensitivity threshold, a transfer learning trigger signal is generated.

[0010] The steps for obtaining real-time data sequence fragments from the preliminary risk situation spectrum, inputting these fragments into a lightweight temporal Transformer network, and calculating the difference between the temporal pattern and the stable pattern are as follows.

[0011] In the preliminary risk situation spectrum, continuous risk situation spectrum data is extracted according to the preset time window length and time step to obtain real-time data sequence fragments;

[0012] Real-time data sequence fragments are input into a lightweight temporal Transformer network, and feature representations are extracted in the encoding layer to generate temporal patterns.

[0013] Historical risk situation spectrum data were collected in batches under normal mine operating conditions, and then the data were input into a lightweight temporal Transformer network to extract temporal pattern features in the coding layer and then statistically averaged to generate stable patterns.

[0014] Substitute the feature vectors of the time series pattern and the stable pattern into the Euclidean distance formula to calculate the difference between the time series pattern and the stable pattern.

[0015] Based on the transfer learning trigger signal, a small batch of data samples is randomly sampled from real-time data sequence fragments to generate a new conditional sample set. The specific steps are as follows.

[0016] Based on the transfer learning trigger signal, small batches of data samples are randomly sampled from real-time data sequence segments according to a preset sampling strategy;

[0017] Each data sample is standardized to generate a standardized data sample. The standardized data sample is then shifted backward by a fixed time step through time shifting to generate a new conditional sample set.

[0018] The disaster prediction model is fine-tuned using a new conditional sample set. Real-time disaster risk indicators are generated by selectively activating connection paths in the deep temporal prediction network by adjusting the mask. The specific steps are as follows.

[0019] A comparative experiment was conducted to evaluate the convergence performance and prediction accuracy of disaster prediction models under different learning rates on a new conditional sample set, and a small step learning rate was obtained.

[0020] Using the new conditional sample set as training data, the parameters of the disaster prediction model are iteratively updated using the gradient descent method under the constraint of a small step learning rate, thereby generating a fine-tuned disaster prediction model.

[0021] By selectively activating connection paths in the deep temporal prediction network using a mask matrix, the disaster prediction model is fine-tuned to re-predict the probability of disaster risk and generate real-time disaster risk indicators.

[0022] As a preferred embodiment of the adaptive transfer learning disaster prediction method under complex geological conditions described in this invention, the steps include: collecting multi-source heterogeneous data from the target mine, cleaning, denoising, and standardizing the multi-source heterogeneous data, and training an initial prediction model using the multi-source heterogeneous data. The specific steps are as follows.

[0023] Collect multi-source heterogeneous data, perform outlier detection and removal on the multi-source heterogeneous data, obtain cleaned multi-source heterogeneous data, and use Kriging interpolation to identify and fill missing values ​​to generate interpolated data. Then, use Z-score standardization to unify the scale of the interpolated data to generate a high-quality dataset.

[0024] The high-quality dataset is divided into training, validation, and test sets, and the initial prediction model is trained to obtain a disaster prediction model.

[0025] As a preferred embodiment of the adaptive transfer learning disaster prediction method under complex geological conditions described in this invention, the method of training the initial prediction model and obtaining the disaster prediction model is achieved by iteratively updating the parameters of the initial prediction model on the training set using backpropagation and gradient descent, evaluating different hyperparameter combinations on the validation set and selecting the optimal performance combination, and evaluating the performance of the initial prediction model under the selected hyperparameters on the test set.

[0026] As a preferred embodiment of the adaptive transfer learning disaster prediction method under complex geological conditions described in this invention, the steps of deploying the disaster prediction model to an edge computing server and establishing a data connection with a real-time sensor network, performing forward computation on new multi-source heterogeneous data, and outputting a preliminary risk situation spectrum are as follows.

[0027] The disaster prediction model is imported into the mine's edge computing server in the form of model files and parameter files. A data transmission link is generated by configuring the data transmission channel between the edge computing server and the real-time sensor network.

[0028] New multi-source heterogeneous data is acquired and transmitted to the disaster prediction model through a data transmission link. The new multi-source heterogeneous data is then nonlinearly mapped and accumulated with time-series features through a temporal neural network layer to generate disaster risk probabilities.

[0029] The disaster risk probability is standardized, and the standardized disaster risk probability is compared with the maximum risk probability within the current time window to output a preliminary risk situation spectrum.

[0030] As a preferred embodiment of the adaptive transfer learning disaster prediction method under complex geological conditions described in this invention, the sensitivity threshold is obtained by analyzing the difference distribution between time-series patterns and stable patterns in historical risk situation data, combined with the actual risk level and expected response time of the mine environment.

[0031] As a preferred embodiment of the adaptive transfer learning disaster prediction method under complex geological conditions described in this invention, the generation of the transfer learning trigger signal refers to comparing the difference degree with the sensitivity threshold. When the difference degree exceeds the sensitivity threshold, the transfer learning mechanism is triggered to adjust the disaster prediction model through transfer learning.

[0032] As a preferred embodiment of the adaptive transfer learning disaster prediction method under complex geological conditions described in this invention, the small step learning rate is obtained by setting candidate learning rates on a new condition sample set, comparing and comprehensively evaluating the convergence speed, convergence stability and prediction accuracy of the disaster prediction model under different candidate learning rates, and selecting the learning rate with the smallest prediction error under the premise of ensuring convergence stability as the small step learning rate.

[0033] The beneficial effects of this invention are as follows: by deploying the disaster prediction model to an edge computing server and connecting it to a real-time sensor network, accurate prediction of real-time risk situation is achieved; by calculating the difference between real-time data and stable patterns, a transfer learning mechanism is triggered to fine-tune the disaster prediction model, thereby improving the model's adaptability and robustness; the model is dynamically adjusted according to changing geological conditions, improving prediction accuracy and reliability, and providing efficient and real-time support for disaster management. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart of an adaptive transfer learning disaster prediction method under complex geological conditions.

[0036] Figure 2 This is a flowchart of the process for collecting multi-source heterogeneous data and training the initial prediction model.

[0037] Figure 3 A flowchart for model deployment and real-time data processing.

[0038] Figure 4 A flowchart for generating transfer learning and real-time risk indicators. Detailed Implementation

[0039] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0040] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0041] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0042] Reference Figures 1-4 This is one embodiment of the present invention, which provides an adaptive transfer learning disaster prediction method under complex geological conditions, comprising the following steps:

[0043] S1. Collect multi-source heterogeneous data from the target mine, clean, denoise, and standardize the multi-source heterogeneous data, train the initial prediction model using the multi-source heterogeneous data, and obtain the disaster prediction model.

[0044] S1.1 Collect multi-source heterogeneous data, perform outlier detection and removal on the multi-source heterogeneous data, obtain cleaned multi-source heterogeneous data, and use Kriging interpolation to identify and fill missing values ​​to generate interpolated data. Use Z-score standardization to unify the scale of the interpolated data to generate a high-quality dataset.

[0045] Furthermore, by collecting multi-source heterogeneous data, the values ​​of each monitoring dimension in the multi-source heterogeneous data are compared with the overall statistical distribution of the multi-source heterogeneous data. Outlier detection methods are used to identify multi-source heterogeneous data samples that significantly deviate from the normal range of variation as outliers, and these outlier samples are removed from the multi-source heterogeneous data to obtain cleaned multi-source heterogeneous data free of outliers. In the cleaned multi-source heterogeneous data, the presence of missing values ​​at each record location is checked. Locations with missing values ​​in the cleaned multi-source heterogeneous data are used as interpolation points for Kriging interpolation, and spatially adjacent observations in the cleaned multi-source heterogeneous data are used as known points for Kriging interpolation. These observations are then calculated... The spatial correlation between the results is used to estimate the missing values ​​at the interpolation points and complete the missing value imputation, thereby generating interpolated data with missing values ​​filled in. In the interpolated data, the statistics of the interpolated data are calculated for each feature dimension, and Z-score standardization is used to map the original value of each feature dimension in the interpolated data to a uniform scale by subtracting the statistics of the corresponding feature dimension and dividing by the statistics of the corresponding feature dimension, so that the values ​​of the interpolated data on different feature dimensions are comparable. Outlier removal, missing value identification and Kriging interpolation imputation and Z-score standardization are applied to multi-source heterogeneous data in sequence to generate a high-quality dataset that meets the training requirements of subsequent disaster prediction models.

[0046] S1.2 Divide the high-quality dataset into training, validation, and test sets. Iterate the parameters of the initial prediction model multiple times on the training set using backpropagation and gradient descent. Evaluate different combinations of hyperparameters on the validation set and select the best-performing combination. Evaluate the performance of the initial prediction model under the selected hyperparameters on the test set to obtain the disaster prediction model.

[0047] Furthermore, for each sample in the high-quality dataset, a partitioning label is set according to a preset partitioning rule. Based on the partitioning label, the high-quality dataset is split into a training set containing samples for parameter learning, a validation set containing samples for generalization performance monitoring, and a test set containing samples for independent testing. Within the training set, samples are divided into batches. The input information recorded in each batch of training samples is input into the forward computation structure of the initial prediction model, obtaining prediction results corresponding one-to-one with the input information recorded in each batch of training samples. By subtracting each batch of prediction results from the target information recorded in each batch of training samples, squaring the resulting differences, summing the squared results, and averaging the summed results, a numerical index is obtained to quantify the prediction error of the initial prediction model on the training set. After obtaining the numerical index, a backpropagation algorithm is used to... The training process calculates the gradient of numerical indicators with respect to the parameters of each layer within the initial prediction model, and updates the parameters of each layer using gradient descent. This process is repeated across multiple training epochs, involving batch sample input, prediction result calculation, prediction error numerical indicator calculation, and gradient descent update operations. This allows the initial prediction model to gradually reduce its prediction error on the training set during consecutive training epochs. On the validation set, the generalization performance of the initial prediction model corresponding to different hyperparameter combinations is monitored, and the optimal hyperparameter combination is selected. On the test set, the prediction model corresponding to the selected hyperparameter combination is set to inference mode, and the disaster risk prediction value is calculated and output batch by batch for the test set samples. The predicted values ​​are compared with the real labels, and the performance indicators are calculated and summarized. These performance indicators are used as the performance evaluation output of the prediction model on the test set, and the prediction model to be evaluated is determined as the disaster prediction model.

[0048] It should be noted that the initial prediction model is a deep time series prediction network, which typically consists of an input layer, a feature fusion layer, a time series modeling layer, a time series neural network layer, and an output layer. The input layer receives multi-source heterogeneous data from the mine. The feature fusion layer fuses data from different sources through nonlinear transformation. The time series modeling layer captures the dependencies between time steps in the time series and learns the dynamic patterns of data changes over time. The time series neural network layer performs nonlinear mapping and time series feature accumulation on the input time series data and outputs the probability of disaster risk.

[0049] S2. Deploy the disaster prediction model to the edge computing server and establish a data connection with the real-time sensor network. Perform forward calculations on new multi-source heterogeneous data and output a preliminary risk situation spectrum.

[0050] S2.1 Import the disaster prediction model into the mine's edge computing server in the form of model files and parameter files, and generate a data transmission link by configuring the data transmission channel between the edge computing server and the real-time sensor network.

[0051] Furthermore, the disaster prediction model is exported from the training environment as model files and parameter files, and transmitted to the mine's edge computing server via wired or wireless network. On the mine's edge computing server, the model and parameter files are imported by setting the storage paths for the model and parameter files and configuring the disaster prediction model loading command, making the disaster prediction model available for execution. A data transmission channel is configured between the mine's edge computing server and the real-time sensor network based on the mine's communication conditions. By assigning communication addresses, port numbers, and communication protocol types to the edge computing server and the real-time sensor network respectively, and setting the data sending direction, data receiving direction, and data refresh cycle in the communication management software, the mine's edge computing server can continuously receive multi-source heterogeneous data collected and uploaded by the real-time sensor network within predetermined time intervals, establishing a stable and usable data transmission link between the mine's edge computing server and the real-time sensor network.

[0052] S2.2 Acquire new multi-source heterogeneous data and transmit the new multi-source heterogeneous data to the disaster prediction model through the data transmission link. Perform nonlinear mapping and time-series feature accumulation on the new multi-source heterogeneous data through the time-series neural network layer to generate disaster risk probability.

[0053] Furthermore, after the disaster prediction model is imported into the mine's edge computing server via model and parameter files, and a data transmission link is established between the mine's edge computing server and the real-time sensor network, the mine's edge computing server receives new multi-source heterogeneous data from the real-time sensor network at predetermined time intervals. This data is then cached in chronological order in the data buffer area of ​​the mine's edge computing server. The new multi-source heterogeneous data is then input into the temporal neural network layer of the disaster prediction model in chronological order. In the temporal neural network layer, a linear transformation is first performed on the feature vector at each time step, and a bias is added. Then, a nonlinear activation function is applied to map the data from the original feature representation to a high-dimensional feature space. The temporal neural network layer accumulates temporal features by passing hidden state vectors between multiple time steps and updating these state vectors, generating a comprehensive feature representation containing time dependencies. This comprehensive feature representation is passed to the output layer of the disaster prediction model, where it undergoes linear transformation, nonlinear activation, and probability normalization operations to generate a disaster risk probability representing the likelihood of a geological disaster occurring.

[0054] S2.3 Standardize the disaster risk probability and compare the standardized disaster risk probability with the maximum risk probability within the current time window to output a preliminary risk situation spectrum.

[0055] Furthermore, the specific steps for standardizing disaster risk probabilities are as follows: A standardization operation is performed on each disaster risk probability value, converting it into a dimensionless value. The standardization operation adjusts the scale of the value based on the mean and standard deviation of the disaster risk probabilities within the current time period, ensuring that disaster risk probabilities obtained from different time periods can be mapped to a unified numerical range. The maximum risk probability within the current time window is the highest-level disaster risk probability value selected by comparing the numerical values ​​among multiple disaster risk probabilities output by the disaster prediction model for each time step within that time window. After standardization, it is ensured that all disaster risk probability values ​​are within the same scale range. The standardized disaster risk probability is compared with the maximum risk probability within the current time window, and the difference between the two is calculated. Based on this difference, a preliminary risk situation spectrum is output.

[0056] It should be noted that the standardized disaster risk probability is compared with the maximum risk probability within the current time window, and the difference between the two is calculated using the following expression:

[0057] ;

[0058] in, Indicates the time within the current time window The corresponding standardized disaster risk probability, This represents the maximum value among the standardized disaster risk probabilities at all times within the current time window. Indicates time Difference measurement.

[0059] S3. Obtain real-time data sequence fragments from the preliminary risk situation spectrum, input the real-time data sequence fragments into a lightweight temporal Transformer network, and calculate the difference between the temporal pattern and the stable pattern. When the difference exceeds the sensitivity threshold, generate a transfer learning trigger signal.

[0060] S3.1. Extract continuous risk situation spectrum data from the preliminary risk situation spectrum according to the preset time window length and time step to obtain real-time data sequence fragments.

[0061] Furthermore, based on the set time window length and time step, continuous time period data are extracted from the preliminary risk situation spectrum to form real-time data sequence segments. Each real-time data sequence segment contains risk situation data within a complete time window, and the time step ensures a fixed time interval between each sequence segment. In this way, the evolution of disaster risk over different time periods can be obtained, providing data support for subsequent time series analysis and forecasting.

[0062] S3.2 Input real-time data sequence fragments into a lightweight temporal Transformer network and extract feature representations in the encoding layer to generate temporal patterns.

[0063] Furthermore, real-time data sequence segments are input sequentially into the encoding layer of a lightweight temporal Transformer network. In this layer, the feature vector of each time step of the real-time data sequence segment undergoes linear transformation and non-linear activation. A self-attention mechanism captures the correlations between time steps, generating a high-dimensional feature representation for each time step. By processing the features from all time steps, a comprehensive temporal feature representation is generated.

[0064] It should be noted that the training process of the lightweight temporal Transformer network is as follows: Sample data is organized chronologically on the training set, and each sample is input into the lightweight temporal Transformer network for forward computation, extracting feature representations through the encoding layer. The output temporal pattern is compared with the target values ​​of the training set samples to calculate the prediction error. The prediction error is then used to calculate the gradient of the network parameters (such as weights and biases) through backpropagation, and these parameters are updated using gradient descent. During training, the network parameters are progressively adjusted to optimize the representation capability of temporal features. After multiple training epochs, the network can effectively learn the temporal patterns in the input data and improve the temporal prediction accuracy. The network performance is then validated on the validation and test sets to select the optimal parameter configuration.

[0065] S3.3. Collect historical risk situation spectrum data in batches under normal mine operating conditions, and input them into a lightweight temporal Transformer network to extract temporal pattern features in the coding layer and then perform statistical averaging to generate stable patterns.

[0066] Furthermore, under normal mine operating conditions, historical risk situation spectrum data is collected in batches and input into a lightweight temporal Transformer network. After the data is input into the encoding layer of the lightweight temporal Transformer network, it undergoes feature extraction processing to generate feature representations for each time step. Then, the feature representations of multiple time steps are statistically averaged to obtain a temporal feature set representing the stable state under normal mine operating conditions; this feature set is the stable mode.

[0067] S3.4 Substitute the feature vectors of the time series mode and the stable mode into the Euclidean distance formula to calculate the difference between the time series mode and the stable mode.

[0068] Furthermore, feature vectors can be extracted from time-series patterns by selecting statistical features of time-series data, such as signal strength, signal-to-noise ratio, bit error rate, and round-trip time, including mean, variance, and extreme values. Feature vectors can be extracted from stable patterns by selecting the average value and fluctuation range of signal strength, signal-to-noise ratio, and bit error rate under stable conditions. The inconsistencies in dimensions between the feature vectors in time-series patterns and stable patterns can be eliminated by normalization or standardization methods. These two feature vectors can be considered as two points for Euclidean distance calculation, and the distance between these two feature vectors, i.e., the degree of difference, can be calculated using the Euclidean distance formula.

[0069] It should be noted that the distance between these two feature vectors is calculated using the Euclidean distance formula, which is expressed as follows:

[0070] ;

[0071] in, The first time sequence pattern 1 eigenvector The first stable mode 1 eigenvector It is the dimension of the feature vector; the calculated Euclidean distance This refers to the degree of difference between the time-series mode and the stable mode.

[0072] S3.5. Compare the difference degree with the sensitivity threshold. When the difference degree exceeds the sensitivity threshold, trigger the transfer learning mechanism to adjust the disaster prediction model through transfer learning and generate a transfer learning trigger signal.

[0073] Furthermore, the difference between the real-time disaster risk probability and the maximum risk probability within the current time window is calculated, representing the degree of deviation between the real-time risk and the maximum risk. By analyzing the distribution of differences between time-series and stable patterns in historical risk situation data, and combining this with the actual risk level and expected response time of the mine environment, a sensitivity threshold is set. The calculated difference is compared with the sensitivity threshold. When the difference exceeds the sensitivity threshold, a transfer learning mechanism is triggered to adjust the disaster prediction model, enabling it to adapt to new operating conditions and improve prediction accuracy and response capability.

[0074] It should be noted that the sensitivity threshold is obtained by analyzing the difference distribution between time-series patterns and stable patterns in historical risk situation data, combined with the actual risk level of the mine environment and the expected response time. The range of the sensitivity threshold can be adjusted according to the specific mine's risk level and the response speed to disaster warnings; for example, in a high-risk mine environment, the sensitivity threshold can be set to a lower value (e.g., 0.2) to trigger the adjustment of the disaster prediction model earlier, while in a low-risk environment, the sensitivity threshold can be appropriately increased (e.g., 0.5) to reduce the frequency of false alarms.

[0075] Transfer learning is a machine learning method that reduces the need to retrain the model for new tasks by applying knowledge learned in one task or environment to another related task or environment. In mine disaster prediction, transfer learning helps the model quickly adapt to changes in the mine environment and operating conditions by transferring model knowledge learned from historical data to new real-time data. When real-time data differs significantly from known stable patterns, transfer learning triggers model adjustments, enabling it to better handle new risk situations and improve prediction accuracy and adaptability.

[0076] S4. Based on the transfer learning trigger signal, randomly sample a small batch of data samples from the real-time data sequence fragments to generate a new conditional sample set.

[0077] S4.1 Based on the transfer learning trigger signal, randomly sample a small batch of data samples from the real-time data sequence fragments according to the preset sampling strategy.

[0078] Furthermore, the time range and index order covered by the real-time data sequence segments are read according to a preset sampling strategy. Control conditions such as the random sampling start position, random sampling interval, and number of time steps in each batch are set in the preset sampling strategy. Then, random sampling positions are generated in the real-time data sequence segments according to the preset sampling strategy. Several time steps are selected from multiple time steps corresponding to the real-time data sequence segments by random number generation. A continuous real-time data sequence segment corresponding to each selected time step is extracted as a small batch data sample. After obtaining a small batch data sample, the preset sampling strategy is repeated to randomly select time steps until a predetermined number of small batch data samples are accumulated. Thus, under the control of the transfer learning trigger signal, a small batch data sample set is formed by random sampling from the real-time data sequence segments according to the preset sampling strategy for the subsequent transfer learning adjustment stage.

[0079] It should be noted that the preset sampling strategy is set by taking into account the time length and time step interval of real-time data sequence segments, the expected trigger frequency of transfer learning trigger signals, and the number of time steps that the small batch of data samples is expected to cover during the method deployment phase, in order to determine the range of random sampling start time step index values ​​and the number of consecutive time steps contained in a single small batch of data samples.

[0080] S4.2. Standardize each data sample to generate a standardized data sample. Shift the standardized data sample backward by a fixed time step through time shifting to generate a new conditional sample set.

[0081] Furthermore, standardization is performed on each data sample. This involves calculating the statistics of each data sample along its corresponding feature dimension, such as the mean and standard deviation. The original value of the data sample along each feature dimension is then subtracted from the mean and divided by the standard deviation, mapping the value of each data sample along each feature dimension to a uniform scale, thus generating standardized data samples with consistent numerical scales. After obtaining the standardized data samples, a time-shift operation is performed on them along the time axis. This involves shifting the feature vector corresponding to each time step in the standardized data samples backward by a fixed time step. During the shift, data sample content that is moved out of the time window is discarded, and standardized data sample content that conforms to the preset strategy is added to the newly entered time position. This ensures that the standardized data samples after the time-shift operation can reflect the time-series relationship that lags behind the original data samples by a fixed time step. All standardized data samples that have undergone standardization and time-shift processing are summarized to generate a new condition sample set for characterizing the constraints of the new operating conditions.

[0082] S5. Fine-tune the disaster prediction model using the new conditional sample set. Generate real-time disaster risk indicators by selectively activating the connection paths in the deep time series prediction network by adjusting the mask.

[0083] S5.1. Conduct comparative experiments on the convergence effect and prediction accuracy of disaster prediction models on new condition sample sets under different learning rates to obtain a small step learning rate.

[0084] Furthermore, several different learning rate combinations were set for the disaster prediction model on the new conditional sample set. Under each learning rate combination, the new conditional sample set was input into the disaster prediction model in batches, and several rounds of gradient descent-based parameter updates were performed on the disaster prediction model. This caused the disaster prediction model to generate a loss curve that changed with the training rounds on the new conditional sample set and output the corresponding disaster risk prediction result. By quantifying the rate of decrease of the loss value with the training rounds, the fluctuation of the loss value during the convergence phase, and the magnitude of the error between the disaster risk prediction result and the target information on the new conditional sample set, the convergence effect index and prediction accuracy index corresponding to each learning rate combination were obtained. The convergence effect index and prediction accuracy index obtained by different learning rate combinations were compared. In the comparison process, the learning rate combination with stable convergence, sufficient loss value decrease and small prediction error was selected first. From this, a small step learning rate that can balance convergence stability and prediction accuracy on the new conditional sample set was determined.

[0085] S5.2 Using the new conditional sample set as training data, the parameters of the disaster prediction model are iteratively updated using the gradient descent method under the constraint of a small step learning rate, thereby generating a fine-tuned disaster prediction model.

[0086] Furthermore, the new conditional sample set is used as the training data input into the disaster prediction model. The new conditional sample set data is divided into batches and sequentially fed into the forward computation structure of the disaster prediction model. The disaster prediction model outputs the disaster risk prediction result corresponding to each batch of new conditional sample set data. By performing a difference operation between the disaster risk prediction result and the target information recorded in the new conditional sample set data, and constructing a loss function based on the difference, the loss value is calculated. This quantifies the prediction error of the disaster prediction model on the new conditional sample set. Under the constraint of a small step learning rate, gradient descent is used according to... The loss value is used to calculate the gradient of each trainable parameter within the disaster prediction model, and the parameters of the disaster prediction model are updated once according to the gradient direction and the small step size learning rate. In multiple training rounds, the batch input of new conditional sample set data, the calculation of disaster risk prediction results, the calculation of loss value, and the gradient descent parameter update operation are repeated to make the loss value of the disaster prediction model on the new conditional sample set gradually decrease and reach the preset convergence criterion. At the end of the training round that meets the convergence criterion, the disaster prediction model after iterative update of gradient descent under the constraints of new conditional sample set data and small step size learning rate is used as the fine-tuned disaster prediction model.

[0087] S5.3. By selectively activating the connection paths in the deep temporal prediction network through the mask matrix, the disaster prediction model is adjusted and fine-tuned to re-predict the probability of disaster risk, generating real-time disaster risk indicators.

[0088] Furthermore, based on the connection structure of each layer of a deep temporal prediction network (typically including an input layer for data reception, a hidden layer for feature fusion, and a core network layer responsible for temporal modeling), a mask matrix element value scheme is set up to correspond one-to-one with each connection path. For example, the sample value 0 and sample value 1 in the mask matrix represent the closed and active states of a connection path in the deep temporal prediction network, respectively. By performing element-wise multiplication of the mask matrix with the weight matrix or intermediate output features of the corresponding connection path in the deep temporal prediction network, selective activation of the connection paths in the deep temporal prediction network is achieved, thus contributing to the change in disaster risk under the current operating conditions. Larger connection paths are retained for participation in the forward computation of the deep time series prediction network, while connection paths that have a greater impact on noise are suppressed. After completing the selective activation of connection paths based on the mask matrix, the disaster prediction model is fine-tuned and a complete forward propagation is performed using new multi-source heterogeneous data input in real time. Under the updated connection path configuration, the deep time series prediction network performs nonlinear mapping and time series feature accumulation on the input time series, outputs the recalculated disaster risk probability, and organizes and labels the recalculated disaster risk probability according to the real-time output format. The organized disaster risk probability is used as a real-time disaster risk indicator to characterize the current mine disaster risk level.

[0089] It should be noted that the mask matrix is ​​obtained by updating the parameters of the disaster prediction model, using the new condition sample set and historical risk situation spectrum data with similar working conditions to statistically quantify the gradient contribution of each connection path in the deep time series prediction network, and dividing the connection path into active and closed states according to the preset contribution threshold, thereby assigning values ​​to the mask matrix elements corresponding to the connection path.

[0090] Small step learning rate is obtained by setting candidate learning rates on a new conditional sample set, comparing and comprehensively evaluating the convergence speed, convergence stability and prediction accuracy of disaster prediction models under different candidate learning rates, and selecting the learning rate with the smallest prediction error while ensuring convergence stability as the small step learning rate.

[0091] In summary, this invention achieves accurate prediction of real-time risk situations by deploying a disaster prediction model to an edge computing server and connecting it to a real-time sensor network; by calculating the difference between real-time data and stable patterns, a transfer learning mechanism is triggered to fine-tune the disaster prediction model, thereby improving the model's adaptability and robustness; and by dynamically adjusting the model according to changing geological conditions, the prediction accuracy and reliability are improved, providing efficient and real-time support for disaster management.

[0092] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An adaptive transfer learning disaster prediction method under complex geological conditions, characterized in that: include, Collect multi-source heterogeneous data from the target mine, clean, denoise and standardize the multi-source heterogeneous data, use the multi-source heterogeneous data to train the initial prediction model, and obtain the disaster prediction model; Disaster prediction models are deployed to edge computing servers and connected to real-time sensor networks to perform forward computation on new multi-source heterogeneous data, outputting a preliminary risk profile, including... The disaster risk probability is standardized, and then compared with the maximum risk probability within the current time window to output a preliminary risk situation spectrum, expressed as: ； in, Indicates the time within the current time window The corresponding standardized disaster risk probability, This represents the maximum value among the standardized disaster risk probabilities at all times within the current time window. Indicates time Difference measurement; The system acquires real-time data sequence fragments from the preliminary risk situation spectrum, inputs these fragments into a lightweight temporal Transformer network, and calculates the difference between the temporal pattern and the stable pattern. When the difference exceeds the sensitivity threshold, a transfer learning trigger signal is generated. The steps for obtaining real-time data sequence fragments from the preliminary risk situation spectrum, inputting these fragments into a lightweight temporal Transformer network, and calculating the difference between the temporal pattern and the stable pattern are as follows. In the preliminary risk situation spectrum, continuous risk situation spectrum data is extracted according to the preset time window length and time step to obtain real-time data sequence fragments; Real-time data sequence fragments are input into a lightweight temporal Transformer network, and feature representations are extracted in the encoding layer to generate temporal patterns. Historical risk situation spectrum data were collected in batches under normal mine operating conditions, and then the data were input into a lightweight temporal Transformer network to extract temporal pattern features in the coding layer and then statistically averaged to generate stable patterns. Substitute the feature vectors of the time series pattern and the stable pattern into the Euclidean distance formula to calculate the difference between the time series pattern and the stable pattern. Based on the transfer learning trigger signal, a small batch of data samples is randomly sampled from real-time data sequence fragments to generate a new conditional sample set. The specific steps are as follows. Based on the transfer learning trigger signal, small batches of data samples are randomly sampled from real-time data sequence segments according to a preset sampling strategy; The preset sampling strategy is set by comprehensively considering the time length and time step interval of real-time data sequence segments, the expected trigger frequency of transfer learning trigger signals, and the number of time steps that the small batch of data samples is expected to cover, to determine the range of random sampling start time step index values ​​and the number of consecutive time steps contained in a single small batch of data samples. Each data sample is standardized to generate a standardized data sample. The standardized data sample is then shifted backward by a fixed time step through time shifting to generate a new conditional sample set. The disaster prediction model is fine-tuned using a new conditional sample set. Real-time disaster risk indicators are generated by selectively activating connection paths in the deep temporal prediction network by adjusting the mask. The specific steps are as follows. A comparative experiment was conducted to evaluate the convergence performance and prediction accuracy of disaster prediction models under different learning rates on a new conditional sample set, and a small step learning rate was obtained. Using the new conditional sample set as training data, the parameters of the disaster prediction model are iteratively updated using the gradient descent method under the constraint of a small step learning rate, thereby generating a fine-tuned disaster prediction model. By selectively activating connection paths in the deep temporal prediction network using a mask matrix, the disaster prediction model is fine-tuned to re-predict the probability of disaster risk and generate a real-time disaster risk index. The mask matrix is ​​obtained by statistically quantifying the gradient contribution of each connection path in the deep time series prediction network after the disaster prediction model completes parameter updates, using the new condition sample set and historical risk situation spectrum data with similar working conditions, and dividing the connection path into active and closed states according to the preset contribution threshold, thereby assigning values ​​to the mask matrix elements corresponding to the connection path one by one.

2. The adaptive transfer learning disaster prediction method under complex geological conditions as described in claim 1, characterized in that: The process involves collecting multi-source heterogeneous data from the target mine, cleaning, denoising, and standardizing the data, and then using this data to train an initial prediction model. The specific steps are as follows: Collect multi-source heterogeneous data, perform outlier detection and removal on the multi-source heterogeneous data, obtain cleaned multi-source heterogeneous data, and use Kriging interpolation to identify and fill missing values ​​to generate interpolated data. Then, use Z-score standardization to unify the scale of the interpolated data to generate a high-quality dataset. The high-quality dataset is divided into training, validation, and test sets, and the initial prediction model is trained to obtain a disaster prediction model.

3. The adaptive transfer learning disaster prediction method under complex geological conditions as described in claim 2, characterized in that: The process of training the initial prediction model to obtain the disaster prediction model involves iteratively updating the parameters of the initial prediction model on the training set using backpropagation and gradient descent, evaluating different combinations of hyperparameters on the validation set and selecting the optimal combination, and evaluating the performance of the initial prediction model under the selected hyperparameters on the test set.

4. The adaptive transfer learning disaster prediction method under complex geological conditions as described in claim 3, characterized in that: The specific steps for deploying the disaster prediction model to an edge computing server and establishing a data connection with a real-time sensor network to perform forward computation on new multi-source heterogeneous data and output a preliminary risk situation spectrum are as follows. The disaster prediction model is imported into the mine's edge computing server in the form of model files and parameter files. A data transmission link is generated by configuring the data transmission channel between the edge computing server and the real-time sensor network. New multi-source heterogeneous data is acquired and transmitted to the disaster prediction model through a data transmission link. The new multi-source heterogeneous data is then nonlinearly mapped and accumulated with time-series features through a temporal neural network layer to generate disaster risk probabilities. The disaster risk probability is standardized, and the standardized disaster risk probability is compared with the maximum risk probability within the current time window to output a preliminary risk situation spectrum.

5. The adaptive transfer learning disaster prediction method under complex geological conditions as described in claim 1, characterized in that: The sensitivity threshold is obtained by analyzing the difference distribution between time-series and stable patterns in historical risk situation data, combined with the actual risk level of the mine environment and the expected response time.

6. The adaptive transfer learning disaster prediction method under complex geological conditions as described in claim 1, characterized in that: The generation of the transfer learning trigger signal refers to comparing the degree of difference with the sensitivity threshold. When the degree of difference exceeds the sensitivity threshold, the transfer learning mechanism is triggered to adjust the disaster prediction model through transfer learning.

7. The adaptive transfer learning disaster prediction method under complex geological conditions as described in claim 1, characterized in that: The small step learning rate is obtained by setting candidate learning rates on a new conditional sample set, comparing and comprehensively evaluating the convergence speed, convergence stability, and prediction accuracy of the disaster prediction model under different candidate learning rates, and selecting the learning rate with the smallest prediction error while ensuring convergence stability as the small step learning rate.

Images