Mine water storage pressure prediction method based on deep learning

Abstract

The invention discloses a deep learning-based mine water storage pressure prediction method, and relates to the technical field of mine water storage pressure prediction. The method comprises the following steps: data acquisition: performing monitoring data acquisition on different nodes of a mine water storage area by using a sensor, and preprocessing the acquired monitoring data to construct a data set; model construction: constructing a deep learning-based mine water storage pressure prediction model for different nodes; model training: inputting data of different nodes into corresponding mine water storage pressure prediction models, and performing cooperative training on the mine water storage pressure prediction models of the nodes by adopting a federated average algorithm to obtain trained mine water storage pressure prediction models; and pressure prediction: using the trained mine water storage pressure prediction model to predict the mine water storage pressure. According to the method, the precision and generalization ability of prediction and classification of the mine water storage pressure can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of prediction of mine water sealing pressure, and particularly to a method for predicting mine water sealing pressure based on deep learning. Background Art

[0002] With the deepening of mine mining activities, parameters such as water sealing pressure, temperature, and humidity inside the mine have an important impact on mine safety and mining efficiency. Traditional mine monitoring systems usually rely on simple sensor data collection and basic analysis algorithms, and there are problems such as insufficient data processing and low model accuracy. Especially when facing multi-dimensional and complex sensor data, the existing technologies are difficult to comprehensively consider the local characteristics, time correlation, and non-stationary distribution of the data, resulting in low accuracy of mine safety prediction and pressure classification.

[0003] The existing technologies often have the following deficiencies: The commonly used weight initialization methods of traditional convolutional neural networks are likely to lead to slow convergence speed or getting stuck in local optimal solutions during the training process. The commonly used optimization algorithms (such as gradient descent method) in existing methods are likely to get stuck in local optimal solutions in high-dimensional parameter spaces and are sensitive to non-stationary distributed data. The traditional learning rate adjustment methods may not work well when facing noise and non-stationary data. Many existing methods fail to perform detailed optimization after preliminary training, resulting in weak ability of the model in detail recognition. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for predicting mine water sealing pressure based on deep learning that can improve the accuracy and generalization ability of mine water sealing pressure prediction classification.

[0005] To solve the above technical problem, the technical solution adopted by the present invention is: A method for predicting mine water sealing pressure based on deep learning, comprising the following steps: Data collection: Using sensors to collect monitoring data of different nodes in the mine water sealing area, and preprocessing the collected monitoring data to construct a data set; Model construction: Adopting a model training architecture of distributed federated learning to construct a deep learning-based mine water sealing pressure prediction model for different nodes; Model training: Inputting the data of different nodes into the corresponding mine water sealing pressure prediction models, and collaboratively training the mine water sealing pressure prediction models of each node by adopting the federated average algorithm to obtain the trained mine water sealing pressure prediction model; Pressure prediction: Using the trained mine water sealing pressure prediction model to predict the mine water sealing pressure.

[0006] The beneficial effects of adopting the above technical solution are as follows: By adopting the model training architecture of distributed federated learning, the present invention realizes the distributed model training and data mining of the multi-node joint model, can avoid the problem of data resource islands of mine water sequestration sensors in the centralized training mode, and aggregates and obtains more effective mine water sequestration sensor data. At the same time, the distributed federated learning architecture can prevent the leakage of local mine water sequestration sensor data of each node. The mine water sequestration sensor data is used locally and the model is trained locally, effectively ensuring the security of the mine water sequestration sensor data. By adopting strategies such as dynamic topology network, quantum tunneling effect, and biotaxis search, it is possible to extract data features more accurately, optimize the model training process, and improve the accuracy and generalization ability of mine water sequestration pressure prediction classification, thereby enhancing the security and reliability of the mine monitoring system. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0008] Figure 1 is the overall flowchart of the method described in the embodiment of the present invention; Figure 2 is the model training architecture diagram of distributed federated learning in the method described in the embodiment of the present invention; Figure 3 is the image diagram of the influence of different normalization methods on feature extraction in the embodiment of the present invention; Figure 4 is the performance comparison curve diagram of different network structures on mine water sequestration data in the embodiment of the present invention; Figure 5 is the influence curve diagram of different initialization methods on the model convergence speed in the embodiment of the present invention; Figure 6 is the robustness comparison curve diagram of different optimization algorithms in a noisy environment in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0009] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0010] Many specific details are set forth in the following description in order to provide a thorough understanding of the present invention, but the present invention may be practiced in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention, so the present invention is not limited by the specific embodiments disclosed below.

[0011] Generally, as Figure 1 shown, an embodiment of the present invention discloses a method for predicting the sealing pressure of mine water based on deep learning, including the following steps: S1, data collection: Use sensors to collect monitoring data of different nodes in the mine water sealing area, and preprocess the collected monitoring data to construct a data set; S2, model construction: Adopt a model training architecture of distributed federated learning to construct a deep learning-based mine water sealing pressure prediction model for different nodes; S3, model training: Input the data of different nodes into the corresponding mine water sealing pressure prediction model, and co-train the mine water sealing pressure prediction models of each node by adopting the federated averaging algorithm to obtain a trained mine water sealing pressure prediction model; S4, pressure prediction: Use the trained mine water sealing pressure prediction model to predict the mine water sealing pressure.

[0012] The above steps will be described in detail below in combination with specific methods: In the federated learning framework, each node does not need to share the mine water sealing sensor data of its local training model. Instead, it trains the local model and sends the updated model to the centralized learning unit for summarization. The model training architecture of distributed federated learning is as Figure 2 shown: In the model training architecture of distributed federated learning, in each round of iteration, each node respectively performs local model training, uploads the trained model parameters to the central server, the central server completes parameter aggregation and update, and sends the updated parameters to each node to start a new round of iteration until the training converges.

[0013] In the model training architecture of distributed federated learning, the federated averaging algorithm is adopted as the model aggregation algorithm. The federated averaging algorithm realizes the co-training of local models through multiple global iterations. Specifically, for each global iteration, let the number of nodes be , the total number of samples it owns be , and the number of samples of the th node be , then the objective function of federated learning is defined as:

[0014]

[0015] In the formula, is the objective function of federated learning, is the model parameter, is the model parameter for the Loss prediction of sensor data for a sample mine water seal as the training loss For the th sample of sensor data for mine water seal The th sample of sensor data for mine water seal and its label. Preferably, the training loss is calculated using the cross-entropy loss function

[0016] Furthermore, for the th node, the objective function of this node is defined as:

[0017] In the formula, is the number of samples of the th node is the objective function of the th node is the distribution of sensor data for mine water seal of the th node

[0018] In one embodiment, taking the parameter update method of the th iteration of the th node as an example, let the parameter gradient of the th node be , then the way to update the model parameters at the th iteration is expressed as:

[0019] In the formula, is the model parameter at the th iteration is the model parameter at the th iteration is the learning rate for the current parameter update is the number of samples of the th node

[0020] Furthermore, the parameter update method of the global model of the central server is expressed as:

[0021] In the formula, is the parameter of the global model of the central server at the th iteration

[0022] Further, repeating the execution of this iterative operation indicates that the global model of the central server and the models of each node are trained. In one embodiment, the preset stopping iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 5000 times.

[0023] For the training data of each node, it mainly comes from the real-time monitoring of various physical quantities such as pressure, temperature, and humidity in the mine water storage environment. The devices for collection include pressure sensors, temperature sensors, humidity sensors, etc. These devices are all installed at different monitoring points in the mine, and the data collected is transmitted to the data processing platform in real time through the Internet of Things technology. The data collection frequency is once per second to ensure that the state changes of the mine water storage system at each time point can be comprehensively reflected. Specifically as follows: Pressure sensor: used to collect the pressure data in the water storage area, and the pressure data is usually recorded in Pa (Pascal); Temperature sensor: used to monitor the temperature in the water storage system, with the unit of degree Celsius (°C); Humidity sensor: used to collect the humidity data in the water storage environment, with the humidity unit of percentage (%); The collected raw data will be transmitted to the cloud data storage platform through wireless network for centralized storage. The data storage format adopts the time series format. Each data record includes the timestamp, pressure value, temperature value, humidity value, and the annotation information of the corresponding device status (such as working, shutdown, etc.).

[0024] The data annotation categories of the present invention mainly include the following categories: Normal state: indicates that the water storage system is in a normal operating state, and the characteristic dimensions such as pressure, temperature, and humidity fluctuate within a reasonable range.

[0025] Abnormal state: indicates that the water storage system has an abnormality, which may include situations such as excessive pressure, too high temperature, too low humidity, etc. This type of data is used to determine the failure or abnormal state of the water storage system in subsequent model training.

[0026] For the constructed mine water storage pressure prediction model, the present invention uses a convolutional neural network as the basic model and makes corresponding improvements, specifically as follows: Process and analyze the mine water storage sensor data, and perform unified scale transformation and missing value filling on the features such as water pressure, temperature, and humidity contained therein, so that in the subsequent steps, it can enter the convolutional layer in a relatively balanced input form for feature extraction. The calculation method is expressed as:

[0027] In the formula, is the The normalization result of a sensor sample, representing the input features after unified processing; is the original sample feature value, representing the measured values such as the water storage pressure measured by the sensor; and respectively represent the minimum and maximum values of this feature in the entire batch of data; is a positive integer.

[0028] After normalizing the data, in order to fully express the local features and time correlation of the mine water storage sensor data, a dynamic topology network is used to adaptively adjust the preliminary feature representation. By analyzing the similarity between samples, the connection weights between network nodes are dynamically changed, and the number of connections is automatically increased or decreased to match the internal structure of the data. The calculation method is expressed as:

[0029] In the formula, is the topological adjustment index of the th sample, representing the relative difference degree of this sample in the entire data distribution; and are respectively the normalized features of the th and the th sensor samples; is the distance metric, representing the size of the difference between samples; is the threshold, representing the scaling scale of the distance information; is the Sigmoid activation function, representing the soft activation of the distance mapping; is the total number of samples, representing the overall range for statistical network topology; and are both positive integers, representing the sample index. Preferably, can be set to 0.5.

[0030] Furthermore, according to the value of , the increase or decrease amplitude of the connection weight of the network nodes is determined. The calculation method is expressed as:

[0031] In the formula, is the weight between node and node at iteration , representing the connection strength after adaptive topology adjustment; is the weight at iteration ; is the learning rate, representing the step size for adjusting the topological structure; is the dynamic threshold, representing the threshold for whether to increase the connection; " " means if If it is too high, increase the weight; otherwise, decrease the weight. Preferably, The value of can be set to 0.1.

[0032] After completing the construction of the dynamic topology network, initialize the initial weights of the convolutional neural network through the quantum tunneling effect to ensure a high weight diversity in the initial stage, thereby accelerating the subsequent learning of the mine water sequestration sensor data. The calculation method is expressed as:

[0033] In the formula, is the initial weight matrix, representing the learnable parameters of the network at the beginning of training; is the number of input features, representing the dimension of the channels where the convolutional kernel acts; is the imaginary unit, representing the characterization of quantum fluctuations in the complex space; is the angular variable.

[0034] Furthermore, implement the calculation of according to the quantum probability distribution, which is expressed as: In the formula,

[0035] In the formula, is the angular variable, representing the weight phase extracted in the quantum random field; is a random variable driven by qubit phase noise, representing the magnitude of the random perturbation of the quantum state; is the average value of the random variable, representing the center of the quantum distribution; is the standard deviation, representing the discrete range of the quantum distribution; is a constant, representing the angular scaling; preferably, The value of can be set to 0.1.

[0036] After completing the initialization, perform forward propagation on the sensor data, and sequentially extract high-dimensional features through the optimized convolutional layer, activation layer, and pooling layer. The calculation method is expressed as:

[0037] In the formula, is the feature map, representing the convolution extraction result of the network for the local receptive field; is the convolutional kernel weight matrix, representing the learnable convolutional parameters; is the convolution operation; is the input data, representing the normalized sensor sampling; is the bias vector, representing the additive correction; is the activation function of the convolutional neural network, representing the non-linear mapping; preferably, The ReLU function can be adopted.

[0038] After each forward propagation is completed, the bio-chemotaxis search is used to optimize the key network parameters to simulate the adaptive migration behavior of organisms in a chemical gradient environment and search for a better solution in the high-dimensional parameter space. The calculation method is expressed as:

[0039] In the formula, is the weight at iteration , representing the new parameters obtained after chemotaxis search; is the weight at iteration ; is the learning rate, representing the step size in chemotaxis search; is the number of search directions, representing multiple chemical gradient directions that can be explored simultaneously; is the th search step size in the direction; is the gradient along this direction; and are both positive integers. Preferably, can be set to 0.01.

[0040] Furthermore, according to the differences in the gradient amplitudes of the mine water sequestration sensor data, the step size is adjusted. The calculation method is expressed as:

[0041] In the formula, is the step size in the current direction, representing the migration speed of organisms in a chemical gradient environment; is the maximum step size, representing the upper limit that the step size may reach; is the adjustment parameter, representing the sensitivity to gradient differences; is the magnitude of the gradient in the current direction; is the gradient threshold, representing the boundary for whether to amplify the step size; is a positive integer, representing the search direction index; Preferably, can be set to 1.0.

[0042] After obtaining the chemotaxis search results, backpropagation is performed to update the weights using the gradient descent method corrected by the quantum tunneling effect, and the jump search ability is enhanced through the non-local quantum potential. The calculation method is expressed as:

[0043] In the formula, is the weight at iteration , representing the parameters after completing the quantum tunneling correction; is the weight at iteration ; is the learning rate, representing the basic step size for updating along the gradient direction; is the weight factor of the gradient term, representing the degree of dependence on the conventional error correction; is the weight factor of the quantum tunneling term, representing the influence of the quantum non - local effect on the update amplitude; is the loss function, representing the error between the predicted classification and the true label; is the gradient of the loss function with respect to the weights; is the quantum tunneling contribution term.

[0044] Furthermore, according to the possible complex distribution of the mine sensor data, a non - linear mapping of the second - order derivative is adopted, and the calculation method is expressed as:

[0045] In the formula, is the correction amount of the quantum tunneling effect on the weights; is the quantum adjustment factor, representing the influence intensity of the quantum fluctuation in the update process; is the arctangent function, representing the non - linear suppression of the second - order derivative; is the second - order derivative of the loss function with respect to the weights; is a positive integer, representing the number of training iterations. Preferably, The value of can be set to 0.3.

[0046] When adopting the gradient descent method with quantum tunneling correction, in order to better cope with the noise and non - stationary distribution in the mine water seal sensor data, the learning rate is dynamically adjusted through the environmental perception adaptive algorithm, and the calculation method is expressed as:

[0047] In the formula, is the learning rate at iteration ; is the learning rate at iteration ; is the adjustment factor, representing the sensitivity to the loss change; is the change of the loss function between two consecutive iterations; is a positive integer. Preferably, is set to 0.1.

[0048] After each training cycle, a part of the mine water seal sensor samples are used as the validation set to measure the accuracy of the network for pressure prediction classification, and the calculation method is expressed as:

[0049] In the formula, is the accuracy rate, representing the proportion of successful classifications; is the number of samples in the validation set; For the The actual category of the samples; The category predicted by the model; is the indicator function; is a positive integer. Preferably, The value can be set to 1000 depending on the size of the dataset.

[0050] After completing the initial training, in order to further improve the detail recognition ability and generalization performance in the prediction and classification of mine water storage pressure, fine-tuning is performed using finer gradients and small batch samples. The calculation method is expressed as:

[0051] In the formula, For iteration The weight of time; For iteration The weight of is the learning rate in the fine-tuning phase; For fine-tuning the dataset The gradient calculated on ; is the selected high confidence mine sample subset; is a positive integer. Preferably, The value of is set to 0.001.

[0052] In one embodiment, by comparing the effects of different normalization methods on feature extraction, the superiority of the method described in this application in processing mine water storage data is verified. Figure 3 The experiment shown compares traditional minimum and maximum normalization, standardization, and decimal scaling methods. The results show that the dynamic normalization method proposed in the method described in this application can better preserve the original distribution characteristics of the sensor data. By combining the dynamic adjustment mechanism of the correlation between features, it provides more balanced input features for subsequent processing, and is significantly better than other methods in classification accuracy, indicating that the unified scale transformation algorithm can adaptively process multi-source heterogeneous sensor data such as water pressure and temperature.

[0053] In this embodiment, a systematic comparison is also made for the network structure design, and the performance of the mine water storage pressure prediction model of the method described in this application is compared with the traditional fully connected network, convolutional neural network and long short-term memory network. Figure 4 The experiment found that the dynamic topology network effectively overcomes the limitations of the fixed network structure in capturing spatiotemporal correlations through the innovative design of real-time analysis of sample similarity and adaptive adjustment of connection weights. It is particularly suitable for processing the local mutation characteristics and time dependence of water storage data in mine environments, and shows faster convergence speed and higher final accuracy during training.

[0054] In this embodiment, the effects of different weight initialization methods on the model convergence performance are also analyzed. Compared with traditional methods such as random initialization, Xavier initialization, and He initialization, the quantum tunneling initialization adopted in this application constructs a more diverse initial weight distribution in the complex space by using the quantum fluctuation effect, as Figure 5 shown. The experimental results show that this method can make the network in a better parameter space at the initial stage of network training, significantly accelerating the model convergence process.

[0055] In this embodiment, the robust performance of different optimization algorithms in a noisy environment is also analyzed. The biotaxis search optimization of the method described in this application is compared with traditional optimizers such as stochastic gradient descent, adaptive moment estimation, and root mean square propagation, as Figure 6 shown. The experiment simulates the common noise interference scenarios in mine sensor data. The results show that by simulating the adaptive migration behavior of organisms in a chemical gradient environment, this technology realizes a more stable parameter search in the high-dimensional parameter space, and can still maintain good classification performance even under high-noise conditions. This anti-interference ability makes it particularly suitable for processing sensor data with noise and outliers collected in the actual mine environment.

[0056] In summary, the method described in this application uses a dynamic topology network to adaptively adjust data, dynamically adjusts the connection weights between network nodes based on the similarity between samples, and automatically increases or decreases the number of connections, which can more accurately capture the local features and temporal correlations in the data and improve the internal structure expression of the data.

[0057] After preliminary training, fine-tuning is performed with more refined gradients and small batches of samples to further improve the detail recognition ability and generalization performance, making the model more accurate in predicting and classifying the mine water sealing pressure.

[0058] Initialize the initial weights of the mine water sealing pressure prediction model through the quantum tunneling effect to ensure that the weights have high diversity in the initial stage, which helps to accelerate subsequent learning and avoid falling into local optimal solutions.

[0059] Combine the biotaxis search method to optimize the key parameters in the mine water sealing pressure prediction model, simulate the adaptive migration behavior of organisms in a chemical gradient environment, and can effectively find better solutions in the high-dimensional parameter space, improving the learning efficiency and performance of the network.

[0060] Combine the quantum tunneling effect to correct the gradient descent method, enhance the jump search ability, and through the correction of the second derivative by using nonlinear mapping, make the network better cope with the noise and non-stationary distribution in the mine water sealing sensor data.

[0061] During the training process, the learning rate is dynamically adjusted using an environmental perception algorithm to cope with the noise and non-stationary characteristics in the data and optimize the model training process.

Claims

1. A method for predicting the sealing pressure of mine water based on deep learning, characterized in that, It includes the following steps: Data collection: Use sensors to collect monitoring data from different nodes in the mine water storage area, and preprocess the collected monitoring data to construct a data set; Model construction: Adopt the model training architecture of distributed federated learning to construct a deep learning-based mine water storage pressure prediction model for different nodes; Model training: Input the data of different nodes into the corresponding mine water storage pressure prediction model, and use the federated average algorithm to co-train the mine water storage pressure prediction models of each node to obtain the trained mine water storage pressure prediction model; Pressure prediction: Use the trained mine water storage pressure prediction model to predict the mine water storage pressure.

2. The method for predicting the sealing pressure of mine water based on deep learning according to claim 1, wherein The sensors in the step of data collection include: Pressure sensor: Used to collect pressure data in the water storage area; Temperature sensor: Used to monitor the temperature in the water storage system; Humidity sensor: Used to collect humidity data in the water storage environment.

3. The method for predicting the pressure of mine water storage based on deep learning according to claim 1, wherein, The data preprocessing in the step of data collection includes: Perform unified scale transformation and missing value filling on the water pressure, temperature, and humidity features included, and the calculation method is expressed as: In the formula, is the normalization result of the th sensor sample, representing the input features after unified processing; is the original sample feature value, representing the measured value measured by the sensor; and respectively represent the minimum and maximum values of this feature in the entire batch of data; is a positive integer.

4. The method for predicting the pressure of mine water sequestration based on deep learning according to claim 1, wherein The data preprocessing also includes: Adopt a dynamic topology network to adaptively adjust the preliminary feature representation. By analyzing the similarity between samples, dynamically change the connection weights between network nodes, and automatically increase or decrease the number of connections to match the internal structure of the data. The calculation method is expressed as: Wherein, is the topological adjustment index of the th sample, representing the relative difference degree of the sample in the whole data distribution; and are respectively the normalized features of the th and the th sensor samples; is the distance metric, representing the magnitude of the difference between samples; is the threshold, representing the scaling scale of the distance information; is the Sigmoid activation function, representing the soft activation of the distance mapping; is the total number of samples, representing the overall range for statistical network topology; and are both positive integers, representing the sample index; According to The value of determines the increase or decrease range of the connection weight of the network node, and the calculation method is expressed as: In the formula, is the weight between node and node at iteration , representing the connection strength after adaptive topology adjustment; is the weight at iteration ; is the learning rate, representing the step size for topology structure adjustment; is the dynamic threshold, representing the threshold for increasing the connection; " means that if is too high, the weight is increased, otherwise it is decreased.

5. The method for predicting the sealing pressure of mine water based on deep learning according to claim 1, wherein Training the model includes the following steps: Perform forward propagation on the sensor data, and sequentially extract high-dimensional features through the optimized convolutional layer, activation layer, and pooling layer. The calculation method is expressed as: wherein, is the feature map, representing the convolution extraction result of the network for the local receptive field; is the convolution kernel weight matrix, representing the learnable convolution parameters; is the convolution operation; is the input data, representing the sensor sampling after normalization; is the bias vector, representing the additive correction; is the activation function of the convolutional neural network, representing the non-linear mapping.

6. The method for predicting the pressure of mine water storage based on deep learning according to claim 5, characterized in that After each forward propagation is completed, use bio-chemotaxis search to optimize the key network parameters to simulate the adaptive migration behavior of organisms in a chemical gradient environment, and search for better solutions in the high-dimensional parameter space. The calculation method is expressed as: In the formula, is the weight at iteration , representing the new parameters obtained after chemotactic search; is the weight at iteration ; is the learning rate, representing the step size in chemotactic search; is the number of search directions, representing multiple chemical gradient directions that can be explored simultaneously; is the th search step size in the direction; is the gradient along this direction; and are both positive integers; Adjust the step size according to the difference in the gradient magnitude of the mine water storage sensor data. The calculation method is expressed as: In the formula, is the step size in the current direction, representing the migration speed of the organism in the chemical gradient environment; is the maximum step size, representing the upper limit that the step size may reach; is the adjustment parameter, representing the sensitivity to the gradient difference; is the gradient magnitude in the current direction; is the gradient threshold, representing the boundary for whether to amplify the step size; is a positive integer, representing the search direction index.

7. The method for predicting the sealing pressure of mine water based on deep learning according to claim 6, wherein After obtaining the chemotaxis search result, perform backpropagation to update the weights using the gradient descent method corrected by the quantum tunneling effect, and enhance the jump search ability through the non-local quantum potential. The calculation method is expressed as: Wherein, is the weight at iteration , representing the parameter after completing the quantum tunneling correction; is the weight at iteration ; is the learning rate, representing the basic step size for updating along the gradient direction; is the weight factor of the gradient term, representing the degree of dependence on the conventional error correction; is the weight factor of the quantum tunneling term, representing the influence of the quantum non-local effect on the update amplitude; is the loss function, representing the error between the predicted classification and the true label; is the gradient of the loss function with respect to the weight; is the quantum tunneling contribution term; Adopt a non-linear mapping of the second derivative according to the possible complex distribution of the mine sensor data. The calculation method is expressed as: In the formula, is the correction amount of the weight due to the quantum tunneling effect; is the quantum adjustment factor, characterizing the influence intensity of quantum fluctuations in the update process; is the arctangent function, characterizing the nonlinear suppression of the second derivative; is the second derivative of the loss function with respect to the weight; is a positive integer, characterizing the number of training iterations.

8. The method for predicting the pressure of mine water storage based on deep learning according to claim 7, wherein When using the gradient descent method corrected by quantum tunneling, dynamically adjust the learning rate through the environment perception adaptive algorithm. The calculation method is expressed as: Wherein, is the learning rate at iteration ; is the learning rate at iteration ; is the adjustment factor, representing the sensitivity to the change in loss; is the change in the loss function between two consecutive iterations; is a positive integer.

9. The method for predicting the sealing pressure of mine water based on deep learning according to claim 1, wherein After each training cycle ends, use some mine water storage sensor samples as a validation set to measure the accuracy of the network's pressure prediction classification. The calculation method is expressed as: In the formula, is the accuracy, which represents the proportion of successful classification; is the number of samples in the validation set; For the The actual category of the samples; The category predicted by the model; is the indicator function; Is a positive integer.

10. The method for predicting the pressure of mine water storage according to claim 1, characterized in that, After completing the preliminary training, perform fine-tuning using finer gradients and small batches of samples. The calculation method is expressed as: Wherein, is the weight at iteration ; is the weight at iteration ; is the learning rate in the fine-tuning stage; is the gradient calculated on the fine-tuning dataset ; is the selected subset of high-confidence mine samples; is a positive integer.

Images