Mine multi-parameter sensing data fusion and analysis method based on machine learning

Abstract

A mine multi-parameter sensing data fusion and analysis method based on machine learning belongs to the field of mine environment monitoring, and comprises the following steps: S1, collecting and manually marking discrete format data and time sequence format data; s2, carrying out feature extraction on the collected mine multi-parameter sensing data in a discrete format, specifically, carrying out feature extraction by adopting a first neural network based on honey pot optimization of chaotic mapping; s3, inputting the collected gas concentration data in the time sequence format into a neural network algorithm optimized based on a star ring principle to perform feature extraction, namely a second neural network; and S4, performing feature cascading on the output of the first neural network and the output of the second neural network, inputting the cascaded features into a classifier for training a classifier model, and specifically, adopting a fractional order neural network algorithm based on dynamic gradient cutting as a classification algorithm for training. According to the invention, data analysis and mine safety assessment are carried out through the mine safety assessment model obtained through training.

Description

Technical Field

[0001] The present invention belongs to the field of mine environment monitoring, and in particular relates to a mine multi-parameter sensor data fusion and analysis method based on machine learning. Background Art

[0002] Due to the complexity and danger of the mine environment, real-time monitoring and accurate assessment of the safety status of the mine are crucial. The monitoring points inside the mine collect key data about the internal environment of the mine by installing various sensors, such as temperature, humidity, wind speed, wind pressure, and gas concentration sensors. Real-time analysis of these data can help identify potential safety risks, so that early warning measures can be taken in advance to ensure the safety of miners and the continuity of mine operations. In the research in this field, the Chinese invention patent with application number CN202410222288.0 proposes a mine ventilation data cleaning method based on machine learning, which collects data from the intelligent ventilation monitoring system through industrial Internet of Things sensing technology, and uses several machine learning models for data cleaning. Specifically, it includes autoregressive models to handle missing values, isolation forest models to detect outliers, and K-means clustering models to reduce redundant data. This method performs well in improving data quality, optimizing ventilation system performance and safety, and is suitable for processing large-scale mine ventilation data sets. The Chinese invention patent with application number CN202311061091.5 proposes a coal mine water hazard prediction system based on data analysis and machine learning technology, including: a data acquisition module for collecting historical accident data of coal mine water hazards and historical environmental data of coal mines to obtain first data; a data processing module for preprocessing the first data to obtain preprocessed data; a prediction model training and evaluation module for training and evaluating a preset coal mine water hazard prediction model based on preprocessed data to obtain a first coal mine water hazard prediction model; a prediction implementation module for predicting coal mine water hazards based on the first coal mine water hazard prediction model. This method can improve the accuracy of coal mine water hazard prediction, reduce false alarms and missed alarms, improve the credibility of prediction results, avoid accidents or reduce the impact of accidents on mine production, and help mine management staff deal with water hazard risks, reduce losses and ensure production safety. The Chinese invention patent with application number CN201610086152.7 proposes a method for evaluating the impact hazard of pre-mining coal seams in mines, including: establishing a three-dimensional digital geological model of the mine area and visualizing it; mastering the spatial occurrence environment of the coal seams; determining the main geological control factors of rock burst; gridding on the horizontal projection of the pre-mining coal seam area, quantifying the main geological control factors of each grid point, and forming a coal seam impact hazard evaluation index system; generating contour line cloud maps of each index; establishing a mine pre-mining coal seam impact hazard evaluation model; evaluating the impact hazard level of each grid point, obtaining the impact hazard level at any position in the pre-mining coal seam area of ​​the mine, and generating and visualizing the impact hazard level cloud map of the pre-mining coal seam area of ​​the mine. By using digital and visualized geological information technology, the impact hazard area can be preliminarily mastered before coal seam mining, and targeted prevention and control measures can be taken in the dangerous area, which has broad application prospects for the safe mining of coal seams in rock burst mines.

[0003] The above technologies are innovative to a certain extent, but there are still the following problems that need to be further solved:

[0004] 1. In the existing technology, feature extraction of neural networks usually cannot effectively and dynamically adjust the network structure to cope with changes in complex data, resulting in deficiencies in anomaly detection and key feature capture.

[0005] 2. Traditional parameter optimization methods lack adaptability to dynamic changes in the training process and cannot guarantee optimization efficiency and stability under multiple parameters and complex environments.

[0006] 3. Existing classifiers are prone to overfitting when processing nonlinear and high-dimensional data. At the same time, they have slow training speeds and weak model generalization capabilities in big data environments. Summary of the invention

[0007] The purpose of the present invention is to propose a mine multi-parameter sensor data fusion and analysis method based on machine learning, and to perform data analysis and mine safety assessment through a trained mine safety assessment model.

[0008] The technical solution adopted by the present invention is:

[0009] The mine multi-parameter sensor data fusion and analysis method based on machine learning includes the following steps:

[0010] S1, collection and annotation of discrete format data and time series format data:

[0011] The main source of data collection is the sensors arranged at various monitoring points inside the mine. These sensors are distributed in multiple key locations in the mine, such as tunnels, working faces and ventilation facilities. Data collection methods include fixed sensor arrangements, mobile collection equipment and automated monitoring systems; the data is divided into two formats, one is discrete format data and the other is time series format data;

[0012] The collected data is uploaded to the cloud platform in real time via wireless transmission, and the discrete format data is stored in structured JSON format, and the time series format data is stored in CSV format; the collected data is manually annotated;

[0013] S2, using the first neural network to extract and train the discrete format data features:

[0014] The collected multi-parameter sensor data of the mine in a discrete format is subjected to feature extraction. The present invention adopts a 5-layer fully connected neural network for feature extraction, which is called the first neural network. In the prior art, some schemes use neural networks for feature extraction. In some neural network structures, problems such as gradient vanishing, gradient explosion, or falling into local optimal solutions may be encountered, which affects the stability of training and the performance of the model. Specifically, the first neural network based on chaotic mapping honeypot optimization is used for feature extraction to extract key features from complex multi-parameter sensor data of the mine. On the basis of the honeypot optimization algorithm, the present invention adopts a dynamic optimization mechanism to allow the network to automatically adjust the position of the honeypot trap in the hidden layer during the training process. The honeypot trap is a specific neuron whose function is to capture abnormal patterns or key features, which may be outliers.

[0015] The training process of the first neural network algorithm based on chaotic map honeypot optimization is as follows:

[0016] (1) Initialize the parameters of the first neural network; (2) The multi-parameter sensor data of the mine is forward propagated through the first neural network; (3) According to the difference between the network output and the expected output, the network parameters are adjusted through chaotic mapping; (4) The parameters are optimized according to the back propagation algorithm; (5) The position of the honeypot trap is dynamically adjusted according to the changes in the captured features to better adapt to the nonlinearity and complexity of the data; (6) After multiple iterations, the feature extraction ability of the first neural network is evaluated;

[0017] S3, using the second neural network to extract and train the time series format data features:

[0018] The collected gas concentration data in a time series format are input into a neural network algorithm optimized based on the star ring principle for feature extraction, which is called the second neural network. Inspired by the fact that star rings in astronomy are stable structures formed by the gravitational interaction of celestial bodies, and their movement follows specific orbital dynamics, the present invention simulates this steady-state motion to achieve the purpose of finding the optimal parameter configuration in the training process of the second neural network.

[0019] Specifically, the training process of the second neural network algorithm optimized based on the star ring principle is as follows:

[0020] (1) Initialize each set of parameters of the second neural network as a star and assign it an initial position and velocity; (2) Simulate the laws of astrophysics and design an interaction mechanics model between stars; (3) Update the velocity and position of each star according to the interaction force between stars; (4) During the parameter update process, detect any star conflicts that may cause instability in the optimization process; (5) After each iteration, adjust the trajectory of the star according to the loss feedback of the network; (6) Repeat the iteration until the preset stop iteration condition is met;

[0021] S4, classifier model training based on the cascade of features of the first neural network and the second neural network:

[0022] The output of the first neural network and the output of the second neural network are feature cascaded, and the cascaded features are input into the classifier to train the classifier model. Specifically, a fractional order neural network algorithm based on dynamic gradient clipping is used as the classification algorithm for training.

[0023] Furthermore, in step S2, the training process of the first neural network algorithm based on chaotic mapping honeypot optimization is as follows:

[0024] S201, initialize the parameters of the first neural network. Specifically, set the weight of the first neural network to be W p , the bias of the first neural network is b p , the weight and bias parameters of the first neural network are initialized using the chaotic mapping function, which is expressed as:

[0025]

[0026] b p (0) =0

[0027] Where W p (0) is the initial weight of the first neural network, b p (0) is the initial bias of the first neural network, represents a normal distribution with mean 0 and variance as the identity matrix, is a normal distribution, ~ indicates that it obeys a specific distribution, and I indicates the identity matrix;

[0028] S202, the multi-parameter sensing data of the mine is forward propagated through the first neural network, and each neuron in the first neural network calculates the output according to the current input and parameter state, which is expressed as:

[0029]

[0030] In the formula, is the output of the l-1th layer of the first neural network; is the weighted input of the lth layer of the first neural network; and are the weight and bias of the first neural network layer l respectively; Sig() is the S i gmi od activation function;

[0031] S203, according to the difference between the network output and the expected output, the network parameters are adjusted through chaotic mapping. Specifically, the learning rate and the honeypot position are adjusted according to the size and complexity of the error. The honeypot position corresponds to the parameter group of the first neural network, that is, one honeypot corresponds to a group of parameters of the first neural network. The process of adjusting the honeypot position is the process of adjusting the change of the parameters of the first neural network, which is expressed as:

[0032]

[0033] In the formula, ΔW p is the change in the weight parameter of the first neural network, Δb p is the change in the bias parameter of the first neural network; L p is the loss function of the first neural network, and are the gradients of the loss function of the first neural network with respect to weights and biases respectively; η p is the learning rate of the first neural network, is the learning rate of the first neural network at the tth iteration, is the learning rate of the first neural network at the t+1th iteration; ChaosMap() is the chaos mapping function;

[0034] The chaotic mapping function introduces periodic and nonlinear adjustment behaviors by utilizing the dependency between the sine function and the current loss value to adapt to different stages of the training process. The way to adjust the learning rate is expressed as:

[0035]

[0036] In the formula, α p and β p It is the parameter for adjusting the sensitivity and periodicity of chaotic response, which is used to control the adjustment amplitude of learning rate; ∈ p is a very small constant used to prevent division by zero errors;

[0037] S204, performing parameter optimization according to the back propagation algorithm, which is different from the gradient descent algorithm, specifically uses multiple potential solutions generated by chaotic mapping to explore the optimal or near-optimal network parameter configuration. The parameter update method of the first neural network is expressed as:

[0038]

[0039] In the formula, is the weight of the first neural network at the t+1th iteration, is the bias of the first neural network at the t+1th iteration, is the weight of the first neural network at the tth iteration, is the bias of the first neural network at the tth iteration;

[0040] S205. Dynamically adjust the position of the honeypot trap according to the changes in the captured features to better adapt to the nonlinearity and complexity of the data. Specifically, the update of the honeypot position depends on the change of the loss function, and the update method is expressed as:

[0041]

[0042] Where Adj() dynamically adjusts the position of the honeypot trap according to the current loss and the current position of the honeypot trap; φ p is the honeypot trap location, is the honeypot trap location of the t+1th iteration, is the honeypot position of the tth iteration;

[0043] The dynamic adjustment function of the honeypot trap location allows the model to better capture and respond to abnormal patterns or changes in important features during training. The implementation is expressed as:

[0044]

[0045] In the formula, γ p and δ p are the adjustment rate and reaction sensitivity, respectively, making the adjustment of the honeypot position more precise and sensitive to the change of the loss function; tanh() is the hyperbolic tangent function;

[0046] S206. After multiple iterations, the feature extraction capability of the first neural network is evaluated. When the loss function of the first neural network is less than a preset threshold, it indicates that the feature extraction capability of the first neural network has reached a level that meets the requirements. Then the iteration is stopped, and the parameter group of the first neural network corresponding to the optimal honeypot trap position of the last iteration is selected as the parameters of the trained first neural network model.

[0047] Furthermore, in step S3, the training process of the second neural network algorithm optimized based on the star ring principle is as follows:

[0048] S301, initialize each set of parameters of the second neural network as a star, and give it an initial position and velocity. The position of the star corresponds to the coordinates of the parameters in the high-dimensional optimization space, and the velocity represents the update speed and direction of the parameters of the second neural network. Suppose the weight of the second neural network is w r , the bias of the second neural network is b r , each star is given an initial position and initial velocity, then the initialization method of position and velocity is expressed as:

[0049]

[0050] In the formula, The stars are given initial positions, The star is given an initial velocity, and is the initial weight and initial bias of the second neural network, Sig() is the Sigmoid activation function, ∈ r is a random variable that conforms to the normal distribution, ⊕ is a vector concatenation operation;

[0051] S302, simulating the laws of dynamics in astrophysics, designing an interactive mechanical model between celestial bodies. Each celestial body is affected by the gravitational force of other celestial bodies. These gravitational forces will be adjusted according to the distance between celestial bodies and the mass of the celestial bodies. The distance between the celestial bodies represents the difference in the parameters of the second neural network, and the mass of the celestial body represents the influence of the parameters of the second neural network. The calculation method of the interactive force between celestial bodies is expressed as:

[0052]

[0053] In the formula, is the interaction force between stars, G is the gravitational constant, m r and m q is the mass of the rth and qth stars, represents the position of the rth star, represents the position of the qth star, ∥∥ represents the L2 norm;

[0054] Among them, the calculation method of the mass of the star is expressed as:

[0055]

[0056] In the formula, β re is the model sensitivity parameter; Var(w r ) is the variance of the weights of the second neural network on the training set, which is used to measure the influence and instability of the parameters; Var() is the variance function; w r is the weight of the second neural network; w q To distinguish from w r The weight of the second neural network, that is, the weight of the second neural network corresponding to the position of the qth star;

[0057] S303, update the speed and position of each star according to the interaction force between the stars. The speed update reflects the optimization direction of the parameters, and the position update directly changes the parameter value. The update method is expressed as:

[0058]

[0059] In the formula, is the position of the star at the t+1th iteration, is the position of the star at the tth iteration; is the velocity of the star at the t+1th iteration, is the velocity of the star at the tth iteration; γ r is the damping coefficient, which is used to simulate environmental resistance; Δt is the star update step size, which represents the iteration interval of the second neural network;

[0060] The calculation method of the damping coefficient is expressed as:

[0061]

[0062] ΔL=|L t -L t-1 |

[0063] In the formula, λ r is the first parameter to adjust the damping sensitivity, κ r It is the second parameter to adjust the damping sensitivity, L t is the loss function of the second neural network at the tth iteration, L t-1 is the loss function of the second neural network at the t-1th iteration. ΔL is the change in the loss function of the second neural network;

[0064] S304. During the parameter update process, any star conflicts that may cause instability in the optimization process are detected. The star conflicts are characterized by excessive parameter update amplitudes. Once a star conflict is detected, the motion strategy of the relevant star will be adjusted to avoid excessive parameter fluctuations. Specifically, the evaluation index for detecting star conflicts is expressed as:

[0065]

[0066] In the formula, δ r Assess indicators for stellar conflict;

[0067] If δ r If it exceeds the preset threshold, it will be adjusted. The adjustment method is expressed as:

[0068]

[0069] In the formula, ← represents the parameter update operation, min() represents the minimum value operation, and v max is the maximum permissible value of the speed to prevent instability caused by excessive parameter updates, K r () is the kinetic energy adjustment function based on the current star's velocity and the gradient of the loss function;

[0070] S305, adopting the orbit correction mechanism, that is, adjusting the motion trajectory of the star according to the loss feedback of the network after each iteration, similar to the fine-tuning of the celestial orbit, to ensure that the optimization direction can be oriented toward the global optimum. Specifically, the orbit correction is performed according to the feedback of the loss function of the second neural network, which is expressed as:

[0071]

[0072] Where η re is the trajectory correction learning rate, is the gradient of the loss function of the second neural network with respect to the position of the rth star;

[0073] S306: Repeat iterative steps S301 to S305 until a preset stop iteration condition is met, indicating that the model training is completed.

[0074] Furthermore, in step S4, the training process of the fractional-order neural network algorithm based on dynamic gradient clipping is as follows:

[0075] S401, initialize the parameters of the fractional-order neural network, and set the weight of the fractional-order neural network to W u , the bias of the fractional-order neural network is b u , the initialization method is expressed as:

[0076]

[0077] In the formula, is the initialization weight matrix of the fractional-order neural network, is the initialization bias vector of the fractional-order neural network, σ u is the standard deviation of the initialization of the fractional-order neural network, randn(d u ,h u ) represents the generation of d from a standard normal distribution u Line u The matrix of columns, d u and h u are the dimensions of the input layer and hidden layer of the fractional-order neural network respectively;

[0078] S402, the cascade features are processed through each layer of the fractional-order neural network, the calculation of each layer output includes the weighted sum of the weight and the input feature, and then a nonlinear transformation is performed through a fractional-order activation function. Specifically, the cascade feature input to the fractional-order neural network is X u , the forward propagation of data is expressed as:

[0079]

[0080] In the formula, is the linear transformation result of the lth layer of the fractional-order neural network, and are the weights and biases of the lth layer of the fractional-order neural network, is the output after the activation function of the fractional-order neural network is processed, Sig fs () is a non-uniform adaptive activation function; Xu is the cascade feature input to the fractional-order neural network;

[0081] The weight of the fractional-order neural network adopts an adaptive dynamic regularization weight matrix processing method, which can adaptively adjust the weight of each layer of the network and enhance the fitting ability of nonlinear data. Its update method is expressed as:

[0082]

[0083] In the formula, λ u is the regularization coefficient of the fractional-order neural network weight, ∈ u is a smoothing term to avoid singular matrices, I is the identity matrix; ∥∥ F represents the Froben i us norm, which ensures the stability of the matrix; For X u The transpose of

[0084] S403, calculating the loss function of the fractional-order neural network, wherein the loss function of the fractional-order neural network is calculated using a cross entropy function with an adaptive gradient penalty, and is expressed as:

[0085]

[0086] Where, L u is the loss function of the fractional-order neural network, y ik is the true label of the i-th sample in the k-th class, is the predicted probability of the i-th sample in the k-th class; is the Frobenius norm of the lth layer weight of the fractional-order neural network, which is used to control the complexity of the model; m pc is the number of samples input to the fractional-order neural network in the current batch, K pc The number of categories of training samples; L pc is the total number of layers of the fractional-order neural network;

[0087] Among them, the regularization parameter of the loss function of the fractional-order neural network is adjusted using an adaptive adjustment mechanism. When the gradient update is too large, the regularization parameter will automatically increase to suppress over-adjustment of the model parameters, thereby preventing the model from overfitting. The adjustment method is expressed as:

[0088]

[0089] In the formula, γ u is the initial regularization strength of the fractional-order neural network, δ u is the adjustment parameter of the fractional-order neural network, represents the L2 norm of the weight gradient of the lth layer of the fractional-order neural network, θ u is the weight influence threshold of the fractional-order neural network;

[0090] S404. According to the derivative of the loss function with respect to the weight of the fractional-order neural network, the weights in the network are updated using the chain rule. The updating method is expressed as:

[0091]

[0092] In the formula, is the gradient of the weight of the lth layer of the fractional-order neural network; is the derivative of the activation function of the fractional-order neural network;

[0093] S405. Adjust the size of the gradient according to the set dynamic clipping strategy to prevent gradient explosion and ensure effective update of the gradient. Specifically, the method of adjusting the gradient by dynamic gradient clipping is expressed as:

[0094]

[0095]

[0096] In the formula, is the gradient of the weight of the lth layer of the pruned fractional-order neural network, Indicates that Each element of is restricted to -θ u and θ u Between u is the gradient clipping threshold, η u is the learning rate of the fractional-order neural network, ← represents the parameter update operation;

[0097] The gradient clipping threshold is set in an adaptive adjustment manner, and the adjustment method is expressed as:

[0098]

[0099] In the formula, κ u is the dynamic adjustment factor for gradient clipping, the L2 norm of the gradient of the fractional-order neural network Used to reflect the overall change amplitude of the gradient;

[0100] At the same time, the learning rate of the fractional-order neural network clipping is dynamically adjusted according to the progress of training, so that the learning rate is larger in the early stage and gradually decreases in the later stage to prevent the model from falling into the local optimum. The adjustment method is expressed as:

[0101]

[0102] Where η u0 is the initial learning rate of the fractional-order neural network, τ u is the adjustment factor of fractional-order neural network pruning, t u is the current iteration number;

[0103] S406. Repeat iterative steps S401 to S405 until the preset stop iteration condition is met, indicating that the model training is completed; use the trained model to process the newly collected samples, realize the analysis of the multi-parameter sensor data of the mine, make corresponding risk warnings for the mine operations, and take necessary safety measures in combination with historical data and expert experience, so as to effectively ensure the safety and efficiency of the mine operations.

[0104] Furthermore, in step S203, the calculation method of the loss function of the first neural network with respect to its gradient is expressed as:

[0105]

[0106] In the formula, o p It is the sample label obtained by outputting the first neural network through the preset Softmax function. is the derivative of the first neural network loss function with respect to the output, and They are the partial derivatives of the output with respect to the weights and biases of the first neural network; is the symbol of partial derivative.

[0107] Furthermore, in step S304, the kinetic energy adjustment function based on the current speed of the star and the gradient of the loss function takes into account the magnitude of the current parameter update speed and the magnitude of the gradient of the loss function to adaptively adjust the update speed of each parameter so that it slows down in the fast descent area to avoid overshoot, and increases power in the area with a smaller gradient to escape from possible local optimality. The calculation method is expressed as:

[0108]

[0109] In the formula, α rs is the kinetic energy regulation influencing factor, is the gradient of the loss function of the second neural network with respect to the position of the rth star, κ ra is the first kinetic energy tuning hyperparameter, ν ra is the second kinetic energy tuning hyperparameter.

[0110] Furthermore, in step S402, the non-uniform adaptive activation function enhances the sensitivity of the activation function in the nonlinear region of the data by utilizing the square term of the linear transformation result. The calculation method is expressed as:

[0111]

[0112] In the formula, α u is the first parameter dynamically adjusted according to the input feature distribution, β u It is the second parameter that is dynamically adjusted according to the input feature distribution.

[0113] Furthermore, in step S404, the calculation method of the derivative of the activation function of the fractional-order neural network is expressed as:

[0114]

[0115] In the formula, Indicates the order is q u Fractional derivative operator of ;

[0116] The fractional derivative operator is calculated as:

[0117]

[0118] Among them, Γ() is the Gamma function; q u is the order of the fractional derivative. Through the fractional derivative operation, the network can process data with different frequency changes more flexibly and enhance the nonlinear expression ability of the model. u is the current iteration number.

[0119] Beneficial effects of the present invention:

[0120] 1. The present invention adopts a honeypot-optimized fully connected neural network (first neural network) based on chaotic mapping, and effectively captures key features and abnormal patterns in complex mine multi-parameter sensor data by dynamically adjusting the position of the honeypot trap in the hidden layer.

[0121] 2. The present invention adopts a neural network (second neural network) optimized based on the star ring principle, and uses the star ring structure in astrophysics to optimize the network parameter configuration, ensuring that the network automatically finds the optimal parameter configuration during the training process and improves the accuracy of feature extraction.

[0122] 3. The present invention cascades the outputs of the first neural network and the second neural network, and inputs the extracted features into a fractional-order neural network based on dynamic gradient clipping for classification. The network improves its adaptability to nonlinear problems through fractional-order derivatives and optimizes the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0123] Figure 1 It is the overall architecture diagram of the present invention. DETAILED DESCRIPTION

[0124] like Figure 1 As shown, the mine multi-parameter sensor data fusion and analysis method based on machine learning includes the following steps:

[0125] S1, collection and annotation of discrete format data and time series format data:

[0126] The main source of data collection is the sensors arranged at various monitoring points inside the mine. These sensors are distributed in multiple key locations in the mine, such as tunnels, working faces and ventilation facilities. Data collection methods include fixed sensor arrangements, mobile collection equipment and automated monitoring systems; the data is divided into two formats, one is discrete format data and the other is time series format data;

[0127] The collected data is uploaded to the cloud platform in real time via wireless transmission, and the discrete format data is stored in a structured JSON format, and the time series format data is stored in a CSV format;

[0128] In this embodiment, the discrete format data includes: wind speed V a , refers to the air flow velocity in the tunnel, in meters per second (m / s); wind pressure P a , refers to the air pressure in the tunnel, the unit is Pascal (Pa); temperature T a , refers to the temperature inside the mine, in degrees Celsius (°C); humidity H a , refers to the air humidity in the mine, the unit is percentage (%); the gas concentration G at a certain moment a , refers to the concentration of harmful gases such as methane and carbon monoxide, in ppm; the status of ventilation facilities S a , refers to the switch status of the damper and window, represented by 0 / 1.

[0129] The time series format data is gas concentration data (data at multiple consecutive moments), ensuring that the monitoring data can reflect the safety status inside the mine. Specifically, the gas concentration is monitored in real time using gas concentration sensors installed at key points in the mine. The key points of the mine include tunnel entrances, working faces, and ventilation systems. The collection interval of the gas concentration data in time series format is 2.5 hours, that is, the gas concentration is collected every 2.5 hours. The entire collection cycle is 0.5 hours, and the collection cycle may span multiple shifts to ensure that different working hours are covered.

[0130] The gas concentration in discrete format is the numerical collection of gas concentration at a certain moment, and the gas concentration in time series format is the numerical collection of gas concentration at multiple consecutive moments; the gas concentration in discrete format is a discrete format value, and the gas concentration in time series format is a time series format vector.

[0131] The collected data is manually labeled; in this embodiment, the labeled categories include: normal working state, abnormal warning state, and dangerous state, a total of 3 categories.

[0132] S2, using the first neural network to extract and train the discrete format data features:

[0133] The collected multi-parameter sensor data of the mine in a discrete format is subjected to feature extraction. The present invention adopts a 5-layer fully connected neural network for feature extraction, which is called the first neural network. In the prior art, some schemes use neural networks for feature extraction. In some neural network structures, problems such as gradient vanishing, gradient explosion, or falling into local optimal solutions may be encountered, which affects the stability of training and the performance of the model. Specifically, the first neural network based on chaotic mapping honeypot optimization is used for feature extraction to extract key features from complex multi-parameter sensor data of the mine. On the basis of the honeypot optimization algorithm, the present invention adopts a dynamic optimization mechanism to allow the network to automatically adjust the position of the honeypot trap in the hidden layer during the training process. The honeypot trap is a specific neuron whose function is to capture abnormal patterns or key features, which may be outliers.

[0134] The training process of the first neural network algorithm based on chaotic map honeypot optimization is as follows:

[0135] S201, initialize the parameters of the first neural network. Specifically, set the weight of the first neural network to be W p , the bias of the first neural network is b p , the weight and bias parameters of the first neural network are initialized using the chaotic mapping function, which is expressed as:

[0136]

[0137] b p (0) =0

[0138] Where W p (0) is the initial weight of the first neural network, b p (0) is the initial bias of the first neural network, represents a normal distribution with mean 0 and variance as the identity matrix, is a normal distribution, ~ indicates that it obeys a specific distribution, and I indicates the identity matrix;

[0139] S202, the multi-parameter sensing data of the mine is forward propagated through the first neural network, and each neuron in the first neural network calculates the output according to the current input and parameter state, which is expressed as:

[0140]

[0141] In the formula, is the output of the l-1th layer of the first neural network; is the weighted input of the lth layer of the first neural network; and are the weight and bias of the first neural network layer l respectively; Sig() is the S i gmi od activation function;

[0142] S203, according to the difference between the network output and the expected output, the network parameters are adjusted through chaotic mapping. Specifically, the learning rate and the honeypot position are adjusted according to the size and complexity of the error. The honeypot position corresponds to the parameter group of the first neural network, that is, one honeypot corresponds to a group of parameters of the first neural network. The process of adjusting the honeypot position is the process of adjusting the change of the parameters of the first neural network, which is expressed as:

[0143]

[0144] In the formula, ΔW p is the change in the weight parameter of the first neural network, Δb p is the change in the bias parameter of the first neural network; L p is the loss function of the first neural network, and are the gradients of the loss function of the first neural network with respect to weights and biases respectively; η p is the learning rate of the first neural network, is the learning rate of the first neural network at the tth iteration, is the learning rate of the first neural network at the t+1th iteration; ChaosMap() is the chaos mapping function;

[0145] In this embodiment, the chaotic mapping function introduces periodic and nonlinear adjustment behaviors by utilizing the dependency between the sine function and the current loss value to adapt to different stages in the training process. The way to adjust the learning rate is expressed as:

[0146]

[0147] In the formula, α p and β p It is the parameter for adjusting the sensitivity and periodicity of chaotic response, which is used to control the adjustment amplitude of learning rate; ∈ pis a very small constant used to prevent division by zero errors; preferably, ∈ p Set to 0.001, α p Set to 2, β p Set to 1.

[0148] The loss function of the first neural network is expressed as follows:

[0149]

[0150] In the formula, o p It is the sample label obtained by outputting the first neural network through the preset Softmax function. is the derivative of the first neural network loss function with respect to the output, and They are the partial derivatives of the output with respect to the weights and biases of the first neural network; is the symbol of partial derivative.

[0151] S204, performing parameter optimization according to the back propagation algorithm, which is different from the gradient descent algorithm, specifically uses multiple potential solutions generated by chaotic mapping to explore the optimal or near-optimal network parameter configuration. The parameter update method of the first neural network is expressed as:

[0152]

[0153] In the formula, is the weight of the first neural network at the t+1th iteration, is the bias of the first neural network at the t+1th iteration, is the weight of the first neural network at the tth iteration, is the bias of the first neural network at the tth iteration;

[0154] S205. Dynamically adjust the position of the honeypot trap according to the changes in the captured features to better adapt to the nonlinearity and complexity of the data. Specifically, the update of the honeypot position depends on the change of the loss function, and the update method is expressed as:

[0155]

[0156] Where Adj() dynamically adjusts the position of the honeypot trap according to the current loss and the current position of the honeypot trap; φ p is the honeypot trap location, is the honeypot trap location of the t+1th iteration, is the honeypot position of the tth iteration;

[0157] The dynamic adjustment function of the honeypot trap location allows the model to better capture and respond to abnormal patterns or changes in important features during training. The implementation is expressed as:

[0158]

[0159] In the formula, γ p and δ p are the adjustment rate and reaction sensitivity, respectively, so that the adjustment of the honeypot position is more precise and sensitive to the change of the loss function; tanh() is the hyperbolic tangent function. Preferably, γ p Set to 0.3, δ p Set to 0.1.

[0160] S206, after multiple iterations, evaluating the feature extraction capability of the first neural network, when the loss function of the first neural network is less than a preset threshold, indicating that the feature extraction capability of the first neural network reaches a level that meets the requirements, then the iteration is stopped, and the parameter group of the first neural network corresponding to the optimal honeypot trap position of the last iteration is selected as the parameters of the trained first neural network model;

[0161] S3, using the second neural network to extract and train the time series format data features:

[0162] The collected gas concentration data in a time series format are input into a neural network algorithm optimized based on the star ring principle for feature extraction, which is called the second neural network. Inspired by the fact that star rings in astronomy are stable structures formed by the gravitational interaction of celestial bodies, and their movement follows specific orbital dynamics, the present invention simulates this steady-state motion to achieve the purpose of finding the optimal parameter configuration in the training process of the second neural network.

[0163] Specifically, the training process of the second neural network algorithm optimized based on the star ring principle is as follows:

[0164] S301, initialize each set of parameters of the second neural network as a star, and give it an initial position and velocity. The position of the star corresponds to the coordinates of the parameters in the high-dimensional optimization space, and the velocity represents the update speed and direction of the parameters of the second neural network. Suppose the weight of the second neural network is w r , the bias of the second neural network is b r , each star is given an initial position and initial velocity, then the initialization method of position and velocity is expressed as:

[0165]

[0166] In the formula, The stars are given initial positions, The star is given an initial velocity, and is the initial weight and initial bias of the second neural network, Sig() is the Sigmoid activation function, ∈ r is a random variable that conforms to the normal distribution, ⊕ is a vector concatenation operation; preferably, ∈ r Set to 0.1.

[0167] S302, simulating the laws of dynamics in astrophysics, designing an interactive mechanical model between celestial bodies. Each celestial body is affected by the gravitational force of other celestial bodies. These gravitational forces will be adjusted according to the distance between celestial bodies and the mass of the celestial bodies. The distance between the celestial bodies represents the difference in the parameters of the second neural network, and the mass of the celestial body represents the influence of the parameters of the second neural network. The calculation method of the interactive force between celestial bodies is expressed as:

[0168]

[0169] In the formula, is the interaction force between stars, G is the gravitational constant, m r and m q is the mass of the rth and qth stars, represents the position of the rth star, represents the position of the qth star, ∥∥ represents the L2 norm;

[0170] Among them, the calculation method of the mass of the star is expressed as:

[0171]

[0172] In the formula, β re is the model sensitivity parameter; Var(w r ) is the variance of the weights of the second neural network on the training set, which is used to measure the influence and instability of the parameters; Var() is the variance function; w r is the weight of the second neural network; w q To distinguish from w r The weight of the second neural network, that is, the weight of the second neural network corresponding to the position of the qth star; preferably, β re Set to 2.

[0173] S303, update the speed and position of each star according to the interaction force between the stars. The speed update reflects the optimization direction of the parameters, and the position update directly changes the parameter value. The update method is expressed as:

[0174]

[0175] In the formula, is the position of the star at the t+1th iteration, is the position of the star at the tth iteration; is the velocity of the star at the t+1th iteration, is the velocity of the star at the tth iteration; γ r is the damping coefficient, used to simulate environmental resistance; Δt is the star update step length, representing the iteration interval of the second neural network; preferably, γ r Set to 0.9.

[0176] The calculation method of the damping coefficient is expressed as:

[0177]

[0178] ΔL=|L t -L t-1 |

[0179] In the formula, λ r is the first parameter to adjust the damping sensitivity, κ r It is the second parameter to adjust the damping sensitivity, L t is the loss function of the second neural network at the tth iteration, L t-1 is the loss function of the second neural network at the t-1th iteration. ΔL is the change in the loss function of the second neural network; preferably, λ r Set to 2, κ r Set to 1.

[0180] S304. During the parameter update process, any star conflicts that may cause instability in the optimization process are detected. The star conflicts are characterized by excessive parameter update amplitudes. Once a star conflict is detected, the motion strategy of the relevant star will be adjusted to avoid excessive parameter fluctuations. Specifically, the evaluation index for detecting star conflicts is expressed as:

[0181]

[0182] In the formula, δ r Assess indicators for stellar conflict;

[0183] If δ r If it exceeds the preset threshold, it will be adjusted. The adjustment method is expressed as:

[0184]

[0185] In the formula, ← represents the parameter update operation, min() represents the minimum value operation, and v max is the maximum permissible value of the speed to prevent instability caused by excessive parameter updates, K r () is a kinetic energy adjustment function based on the current star's velocity and the gradient of the loss function; preferably, v max Set to 0.9.

[0186] The kinetic energy regulation function based on the current speed of the star and the gradient of the loss function takes into account the size of the current parameter update speed and the amplitude of the loss function gradient to adaptively adjust the update speed of each parameter so that it slows down in the fast descent area to avoid overshoot, and increases the power in the area with smaller gradient to escape the possible local optimum. The calculation method is expressed as:

[0187]

[0188] In the formula, α rs is the kinetic energy regulation influencing factor, is the gradient of the loss function of the second neural network with respect to the position of the rth star, κ ra is the first kinetic energy tuning hyperparameter, ν ra is the second kinetic energy adjustment hyperparameter. Preferably, α rs Set to 0.01, κ ra Set to 2, ν ra Set to 1.

[0189] S305, adopting the orbit correction mechanism, that is, adjusting the motion trajectory of the star according to the loss feedback of the network after each iteration, similar to the fine-tuning of the celestial orbit, to ensure that the optimization direction can be oriented toward the global optimum. Specifically, the orbit correction is performed according to the feedback of the loss function of the second neural network, which is expressed as:

[0190]

[0191] Where η re is the trajectory correction learning rate, is the gradient of the loss function of the second neural network with respect to the position of the rth star; preferably, η re Set to 0.01.

[0192] S306. Repeat iterative steps S301 to S305 until a preset stop iteration condition is met, indicating that the model training is completed. In this embodiment, the preset stop iteration condition is to reach a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0193] S4, classifier model training based on the cascade of features of the first neural network and the second neural network:

[0194] The output of the first neural network and the output of the second neural network are feature cascaded, and the cascaded features are input into the classifier to train the classifier model. Specifically, a fractional-order neural network algorithm based on dynamic gradient clipping is used as the classification algorithm for training; the fractional-order neural network algorithm structure consists of 4 layers of fractional-order neuron layers, and each neuron uses fractional-order derivatives when processing the input cascade features to enhance the model's adaptability to nonlinear problems. The present invention dynamically adjusts the gradient penalty factor according to the update frequency and amplitude of each layer of network weights, thereby optimizing the weight adjustment during the training process, ensuring faster convergence speed and higher model generalization ability, and avoiding overfitting or instability during the training process.

[0195] Specifically, the training process of the fractional-order neural network algorithm based on dynamic gradient clipping is as follows:

[0196] S401, initialize the parameters of the fractional-order neural network, and set the weight of the fractional-order neural network to W u , the bias of the fractional-order neural network is b u , the initialization method is expressed as:

[0197]

[0198] In the formula, is the initialization weight matrix of the fractional-order neural network, is the initialization bias vector of the fractional-order neural network, σ u is the standard deviation of the initialization of the fractional-order neural network, randn(d u ,h u ) represents the generation of d from a standard normal distribution u Line u The matrix of columns, d u and h u are the dimensions of the input layer and hidden layer of the fractional-order neural network respectively;

[0199] S402, the cascade features are processed through each layer of the fractional-order neural network, the calculation of each layer output includes the weighted sum of the weight and the input feature, and then a nonlinear transformation is performed through a fractional-order activation function. Specifically, the cascade feature input to the fractional-order neural network is X u , the forward propagation of data is expressed as:

[0200]

[0201] In the formula, is the linear transformation result of the lth layer of the fractional-order neural network, and are the weights and biases of the lth layer of the fractional-order neural network, is the output after the activation function of the fractional-order neural network is processed, Sig fs () is a non-uniform adaptive activation function; Xu is the cascade feature input to the fractional-order neural network;

[0202] The weight of the fractional-order neural network adopts an adaptive dynamic regularization weight matrix processing method, which can adaptively adjust the weight of each layer of the network and enhance the fitting ability of nonlinear data. Its update method is expressed as:

[0203]

[0204] In the formula, λ u is the regularization coefficient of the fractional-order neural network weight, ∈ u is a smoothing term to avoid singular matrices, I is the identity matrix; ∥∥ F represents the Froben i us norm, which ensures the stability of the matrix; For X u The transpose of u Set to 0.2.

[0205] The non-uniform adaptive activation function enhances the sensitivity of the activation function in the nonlinear region of the data by utilizing the square term of the linear transformation result. The calculation method is expressed as:

[0206]

[0207] In the formula, α u is the first parameter dynamically adjusted according to the input feature distribution, β u is a second parameter dynamically adjusted according to the input feature distribution. Preferably, α u Set to 0.1, β u Set to 0.3.

[0208] S403, calculating the loss function of the fractional-order neural network, wherein the loss function of the fractional-order neural network is calculated using a cross entropy function with an adaptive gradient penalty, and is expressed as:

[0209]

[0210] Where, L u is the loss function of the fractional-order neural network, y ik is the true label of the i-th sample in the k-th class, is the predicted probability of the i-th sample in the k-th class; is the Frobenius norm of the lth layer weight of the fractional-order neural network, which is used to control the complexity of the model; m pc is the number of samples input to the fractional-order neural network in the current batch, K pcThe number of categories of training samples; L pc is the total number of layers of the fractional-order neural network;

[0211] Among them, the regularization parameter of the loss function of the fractional-order neural network is adjusted using an adaptive adjustment mechanism. When the gradient update is too large, the regularization parameter will automatically increase to suppress over-adjustment of the model parameters, thereby preventing the model from overfitting. The adjustment method is expressed as:

[0212]

[0213] In the formula, γ u is the initial regularization strength of the fractional-order neural network, δ u is the adjustment parameter of the fractional-order neural network, represents the L2 norm of the weight gradient of the lth layer of the fractional-order neural network, θ u is the weight influence threshold of the fractional-order neural network; preferably, γ u Set to 2, δ u Set to 2, θ u Set to 0.5.

[0214] S404. According to the derivative of the loss function with respect to the weight of the fractional-order neural network, the weights in the network are updated using the chain rule. The updating method is expressed as:

[0215]

[0216] In the formula, is the gradient of the weight of the lth layer of the fractional-order neural network; is the derivative of the activation function of the fractional-order neural network;

[0217] The calculation method of the derivative of the activation function of the fractional-order neural network is expressed as:

[0218]

[0219] In the formula, Indicates the order is q u The fractional derivative operator for .

[0220] The fractional derivative operator is calculated as:

[0221]

[0222] Among them, Γ() is the Gamma function; q u is the order of the fractional derivative. Through the fractional derivative operation, the network can process data with different frequency changes more flexibly and enhance the nonlinear expression ability of the model. u is the current iteration number.

[0223] S405. Adjust the size of the gradient according to the set dynamic clipping strategy to prevent gradient explosion and ensure effective update of the gradient. Specifically, the method of adjusting the gradient by dynamic gradient clipping is expressed as:

[0224]

[0225] In the formula, is the gradient of the weight of the lth layer of the pruned fractional-order neural network, Indicates that Each element of is restricted to -θ u and θ u Between u is the gradient clipping threshold, η u is the learning rate of the fractional-order neural network, ← represents the parameter update operation;

[0226] The gradient clipping threshold is set in an adaptive adjustment manner, and the adjustment method is expressed as:

[0227]

[0228] In the formula, κ u is the dynamic adjustment factor for gradient clipping, the L2 norm of the gradient of the fractional-order neural network It is used to reflect the overall change amplitude of the gradient; preferably, κ u Set to 0.1.

[0229] At the same time, the learning rate of the fractional-order neural network clipping is dynamically adjusted according to the progress of training, so that the learning rate is larger in the early stage and gradually decreases in the later stage to prevent the model from falling into the local optimum. The adjustment method is expressed as:

[0230]

[0231] Where η u0 is the initial learning rate of the fractional-order neural network, τ u is the adjustment factor of fractional-order neural network pruning, t u is the current iteration number; preferably, η u0 Set to 0.01, τ u Set to 2.

[0232] S406. Repeat iterative steps S401 to S405 until a preset stop iteration condition is met, indicating that the model training is completed; in this embodiment, the preset stop iteration condition is to reach a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0233] The trained model is used to process newly collected samples, analyze the multi-parameter sensor data of the mine, issue corresponding risk warnings for mine operations, and take necessary safety measures based on historical data and expert experience, thereby effectively ensuring the safety and efficiency of mine operations.

Claims

1. A mine multi-parameter sensor data fusion and analysis method based on machine learning, characterized in that: The following steps are involved: S1, collection and annotation of discrete format data and time series format data: The main source of data collection is the sensors arranged at each monitoring point inside the mine. The data is divided into two formats, one is discrete format data and the other is time series format data; The collected data is uploaded to the cloud platform in real time via wireless transmission, and the discrete format data is stored in structured JSON format, and the time series format data is stored in CSV format; at the same time, the collected data is manually annotated; S2, using the first neural network to extract and train the discrete format data features: The collected mine multi-parameter sensor data in discrete format is subjected to feature extraction. Specifically, the first neural network optimized by honeypot based on chaotic mapping is used for feature extraction. The training process of the first neural network algorithm is as follows: (1) Initialize the parameters of the first neural network; (2) The multi-parameter sensor data of the mine is forward propagated through the first neural network; (3) According to the difference between the network output and the expected output, the network parameters are adjusted through chaotic mapping; (4) The parameters are optimized according to the back propagation algorithm; (5) The position of the honeypot trap is dynamically adjusted according to the changes in the captured features to better adapt to the nonlinearity and complexity of the data; (6) After multiple iterations, the feature extraction ability of the first neural network is evaluated; S3, using the second neural network to extract and train the time series format data features: The collected gas concentration data in time series format is input into the neural network algorithm optimized based on the star ring principle for feature extraction, which is called the second neural network. Specifically, the training process of the second neural network algorithm optimized based on the star ring principle is as follows: (1) Initialize each set of parameters of the second neural network as a star and assign it an initial position and velocity; (2) Simulate the laws of astrophysics and design an interaction mechanics model between stars; (3) Update the velocity and position of each star according to the interaction force between stars; (4) During the parameter update process, detect any star conflicts that may cause instability in the optimization process; (5) After each iteration, adjust the trajectory of the star according to the loss feedback of the network; (6) Repeat the iteration until the preset stop iteration condition is met; S4, classifier model training based on the cascade of features of the first neural network and the second neural network: The output of the first neural network and the output of the second neural network are feature cascaded, and the cascaded features are input into the classifier to train the classifier model. Specifically, a fractional order neural network algorithm based on dynamic gradient clipping is used as the classification algorithm for training.

2. The mine multi-parameter sensor data fusion and analysis method based on machine learning according to claim 1, characterized in that: In step S2, the training process of the first neural network algorithm based on chaotic mapping honeypot optimization is as follows: S201, initialize the parameters of the first neural network. Specifically, set the weight of the first neural network to be W p , the bias of the first neural network is b p , the weight and bias parameters of the first neural network are initialized using the chaotic mapping function, which is expressed as: b p (0) ＝0 Where W p (0) is the initial weight of the first neural network, b p (0) is the initial bias of the first neural network, represents a normal distribution with mean 0 and variance as the identity matrix, is a normal distribution, ~ indicates that it obeys a specific distribution, and I indicates the unit matrix; S202, the multi-parameter sensing data of the mine is forward propagated through the first neural network, and each neuron in the first neural network calculates the output according to the current input and parameter state, which is expressed as: In the formula, is the output of the l-1th layer of the first neural network; is the weighted input of the lth layer of the first neural network; and are the weight and bias of the first neural network layer l respectively; Sig() is the Sigmiod activation function; S203, according to the difference between the network output and the expected output, the network parameters are adjusted through chaotic mapping. Specifically, the learning rate and the honeypot position are adjusted according to the size and complexity of the error. The honeypot position corresponds to the parameter group of the first neural network, that is, one honeypot corresponds to a group of parameters of the first neural network. The process of adjusting the honeypot position is the process of adjusting the change of the parameters of the first neural network, which is expressed as: In the formula, ΔW p is the change in the weight parameter of the first neural network, Δb p is the change in the bias parameter of the first neural network; L p is the loss function of the first neural network, and They are the gradients of the loss function of the first neural network with respect to weights and biases respectively; η p is the learning rate of the first neural network, is the learning rate of the first neural network at the tth iteration, is the learning rate of the first neural network at the t+1th iteration; ChaosMap() is the chaos mapping function; The chaotic mapping function introduces periodic and nonlinear adjustment behaviors by utilizing the dependency between the sine function and the current loss value to adapt to different stages of the training process. The way to adjust the learning rate is expressed as: In the formula, α p and β p It is the parameter for adjusting the sensitivity and periodicity of chaotic response, which is used to control the adjustment amplitude of learning rate; ∈ p is a very small constant used to prevent division by zero errors; S204, performing parameter optimization according to the back propagation algorithm, specifically using multiple potential solutions generated by chaotic mapping to explore the optimal or near-optimal network parameter configuration, the parameter update method of the first neural network is expressed as: In the formula, is the weight of the first neural network at the t+1th iteration, is the bias of the first neural network at the t+1th iteration, is the weight of the first neural network at the tth iteration, is the bias of the first neural network at the tth iteration; S205. Dynamically adjust the position of the honeypot trap according to the changes in the captured features to better adapt to the nonlinearity and complexity of the data. Specifically, the update of the honeypot position depends on the change of the loss function, and the update method is expressed as: Where Adj() dynamically adjusts the position of the honeypot trap according to the current loss and the current position of the honeypot trap; φ p is the honeypot trap location, is the honeypot trap location of the t+1th iteration, is the honeypot position of the tth iteration; The dynamic adjustment function of the honeypot trap location allows the model to better capture and respond to abnormal patterns or changes in important features during training. The implementation is expressed as: In the formula, γ p and δ p are the adjustment rate and reaction sensitivity, respectively, making the adjustment of the honeypot position more precise and sensitive to the change of the loss function; tanh() is the hyperbolic tangent function; S206. After multiple iterations, the feature extraction capability of the first neural network is evaluated. When the loss function of the first neural network is less than a preset threshold, it indicates that the feature extraction capability of the first neural network has reached a level that meets the requirements. Then the iteration is stopped, and the parameter group of the first neural network corresponding to the optimal honeypot trap position of the last iteration is selected as the parameters of the trained first neural network model.

3. The mine multi-parameter sensor data fusion and analysis method based on machine learning according to claim 1, characterized in that: In step S3, the training process of the second neural network algorithm optimized based on the star ring principle is as follows: S301, initialize each set of parameters of the second neural network as a star, and give it an initial position and velocity. The position of the star corresponds to the coordinates of the parameters in the high-dimensional optimization space, and the velocity represents the update speed and direction of the parameters of the second neural network. Suppose the weight of the second neural network is w r , the bias of the second neural network is b r , each star is given an initial position and initial velocity, then the initialization method of position and velocity is expressed as: In the formula, The stars are given initial positions, The star is given an initial velocity, and is the initial weight and initial bias of the second neural network, Sig() is the Sigmoid activation function, ∈ r is a random variable that follows a normal distribution. It is a vector concatenation operation; S302. Simulating the dynamics laws in astrophysics, designing an interaction mechanics model between celestial bodies, the calculation method of the interaction force between celestial bodies is expressed as: In the formula, is the interaction force between stars, G is the gravitational constant, m r and m q is the mass of the rth and qth stars, represents the position of the rth star, represents the position of the qth star, ∥∥ represents the L2 norm; Among them, the calculation method of the mass of the star is expressed as: In the formula, β re is the model sensitivity parameter; Var(w r ) is the variance of the weights of the second neural network on the training set, which is used to measure the influence and instability of the parameters; Var() is the variance function; w r is the weight of the second neural network; w q To distinguish from w r The weight of the second neural network, that is, the weight of the second neural network corresponding to the position of the qth star; S303, update the speed and position of each star according to the interaction force between the stars. The speed update reflects the optimization direction of the parameters, and the position update directly changes the parameter value. The update method is expressed as: In the formula, is the position of the star at the t+1th iteration, is the position of the star at the tth iteration; is the velocity of the star at the t+1th iteration, is the velocity of the star at the tth iteration; γ r is the damping coefficient, which is used to simulate environmental resistance; Δt is the star update step size, which represents the iteration interval of the second neural network; The calculation method of the damping coefficient is expressed as: ΔL=|L t -L t-1 | In the formula, λ r is the first parameter to adjust the damping sensitivity, κ r It is the second parameter to adjust the damping sensitivity, L t is the loss function of the second neural network at the tth iteration, L t-1 is the loss function of the second neural network at the t-1th iteration, and ΔL is the change in the loss function of the second neural network; S304. During the parameter update process, any star conflicts that may cause instability in the optimization process are detected. The star conflicts are characterized by excessive parameter update amplitudes. Once a star conflict is detected, the motion strategy of the relevant star will be adjusted to avoid excessive parameter fluctuations. Specifically, the evaluation index for detecting star conflicts is expressed as: In the formula, δ r Assess indicators for stellar conflict; If δ r If it exceeds the preset threshold, it will be adjusted. The adjustment method is expressed as: In the formula, ← represents the parameter update operation, min() represents the minimum value operation, and v max is the maximum permissible value of the speed to prevent instability caused by excessive parameter updates, K r () is the kinetic energy adjustment function based on the current star's velocity and the gradient of the loss function; S305, adopting the orbit correction mechanism, that is, adjusting the motion trajectory of the star according to the loss feedback of the network after each iteration to ensure that the optimization direction can be oriented toward the global optimum. Specifically, the orbit correction is performed according to the feedback of the loss function of the second neural network, which is expressed as: Where η re is the trajectory correction learning rate, is the gradient of the loss function of the second neural network with respect to the position of the rth star; S306: Repeat iterative steps S301 to S305 until a preset stop iteration condition is met, indicating that the model training is completed.

4. The mine multi-parameter sensor data fusion and analysis method based on machine learning according to claim 1, characterized in that: In step S4, the training process of the fractional-order neural network algorithm based on dynamic gradient clipping is as follows: S401, initialize the parameters of the fractional-order neural network, and set the weight of the fractional-order neural network to W u , the bias of the fractional-order neural network is b u , the initialization method is expressed as: In the formula, is the initialization weight matrix of the fractional-order neural network, is the initialization bias vector of the fractional-order neural network, σ u is the standard deviation of the initialization of the fractional-order neural network, randn(d u ,h u ) represents the generation of d from a standard normal distribution u Line u The matrix of columns, d u and h u are the dimensions of the input layer and hidden layer of the fractional-order neural network respectively; S402, the cascade features are processed through each layer of the fractional-order neural network, the calculation of each layer output includes the weighted sum of the weight and the input feature, and then a nonlinear transformation is performed through a fractional-order activation function. Specifically, the cascade feature input to the fractional-order neural network is X u , the forward propagation of data is expressed as: In the formula, is the linear transformation result of the lth layer of the fractional-order neural network, and are the weights and biases of the lth layer of the fractional-order neural network, is the output after the activation function of the fractional-order neural network is processed, Sig fs () is a non-uniform adaptive activation function; Xu is the cascade feature input to the fractional-order neural network; The weight of the fractional-order neural network adopts an adaptive dynamic regularization weight matrix processing method, which can adaptively adjust the weight of each layer of the network and enhance the fitting ability of nonlinear data. Its update method is expressed as: In the formula, λ u is the regularization coefficient of the fractional-order neural network weight, ∈ u is a smoothing term to avoid singular matrices, I is the identity matrix; ∥∥ F represents the Frobenius norm, which ensures the stability of the matrix; For X u The transpose of S403, calculating the loss function of the fractional-order neural network, wherein the loss function of the fractional-order neural network is calculated using a cross entropy function with an adaptive gradient penalty, and is expressed as: Where, L u is the loss function of the fractional-order neural network, y ik is the true label of the i-th sample in the k-th class, is the predicted probability of the i-th sample in the k-th class; is the Frobenius norm of the lth layer weight of the fractional-order neural network, which is used to control the model complexity; m pc is the number of samples input to the fractional-order neural network in the current batch, K pc The number of categories of training samples; L pc is the total number of layers of the fractional-order neural network; Among them, the regularization parameter of the loss function of the fractional-order neural network is adjusted using an adaptive adjustment mechanism. When the gradient update is too large, the regularization parameter will automatically increase to suppress over-adjustment of the model parameters, thereby preventing the model from overfitting. The adjustment method is expressed as: In the formula, γ u is the initial regularization strength of the fractional-order neural network, δ u is the adjustment parameter of the fractional-order neural network, represents the L2 norm of the weight gradient of the lth layer of the fractional-order neural network, θ u is the weight influence threshold of the fractional-order neural network; S404. According to the derivative of the loss function with respect to the weight of the fractional-order neural network, the weights in the network are updated using the chain rule. The updating method is expressed as: In the formula, is the gradient of the weight of the lth layer of the fractional-order neural network; is the derivative of the activation function of the fractional-order neural network; S405. Adjust the size of the gradient according to the set dynamic clipping strategy to prevent gradient explosion and ensure effective update of the gradient. Specifically, the method of adjusting the gradient by dynamic gradient clipping is expressed as: In the formula, is the gradient of the weight of the lth layer of the pruned fractional-order neural network, Indicates that Each element of is restricted to -θ u and θ u Between u is the gradient clipping threshold, η u is the learning rate of the fractional-order neural network, ← represents the parameter update operation; The gradient clipping threshold is set in an adaptive adjustment manner, and the adjustment method is expressed as: In the formula, κ u is the dynamic adjustment factor for gradient clipping, the L2 norm of the gradient of the fractional-order neural network Used to reflect the overall change amplitude of the gradient; At the same time, the learning rate of the fractional-order neural network clipping is dynamically adjusted according to the progress of training, so that the learning rate is larger in the early stage and gradually decreases in the later stage to prevent the model from falling into the local optimum. The adjustment method is expressed as: Where η u0 is the initial learning rate of the fractional-order neural network, τ u is the adjustment factor of fractional-order neural network pruning, t u is the current iteration number; S406: Repeat iterative steps S401 to S405 until a preset stop iteration condition is met, indicating that the model training is completed.

5. The mine multi-parameter sensor data fusion and analysis method based on machine learning as claimed in claim 2, characterized in that: In step S203, the calculation method of the loss function of the first neural network with respect to its gradient is expressed as: In the formula, o p It is the sample label obtained by outputting the first neural network through the preset Softmax function. is the derivative of the first neural network loss function with respect to the output, and They are the partial derivatives of the output with respect to the weights and biases of the first neural network; is the symbol of partial derivative.

6. The mine multi-parameter sensor data fusion and analysis method based on machine learning as claimed in claim 3, characterized in that: In step S304, the kinetic energy adjustment function based on the current star speed and the loss function gradient considers the current parameter update speed and the magnitude of the loss function gradient to adaptively adjust the update speed of each parameter so that it slows down in the fast descent area to avoid overshoot, and increases power in the area with smaller gradient to escape from possible local optimum. The calculation method is expressed as: In the formula, α rs is the kinetic energy regulation influencing factor, is the gradient of the loss function of the second neural network with respect to the position of the rth star, κ ra is the first kinetic energy tuning hyperparameter, ν ra is the second kinetic energy tuning hyperparameter.

7. The mine multi-parameter sensor data fusion and analysis method based on machine learning as claimed in claim 4, characterized in that: In step S402, the non-uniform adaptive activation function enhances the sensitivity of the activation function in the nonlinear region of the data by utilizing the square term of the linear transformation result. The calculation method is expressed as: In the formula, α u is the first parameter dynamically adjusted according to the input feature distribution, β u It is the second parameter that is dynamically adjusted according to the input feature distribution.

8. The mine multi-parameter sensor data fusion and analysis method based on machine learning as claimed in claim 4, characterized in that: In step S404, the calculation method of the derivative of the activation function of the fractional-order neural network is expressed as: In the formula, Indicates the order is q u Fractional derivative operator of ; The fractional derivative operator is calculated as: Among them, Γ() is the Gamma function; q u is the order of the fractional derivative; t u is the current iteration number.

Images