A method for extracting time sequence features of coal mine fire by fusing adaptive multi-granularity information particles and transformers

Abstract

The application discloses a kind of Transform time sequence feature extraction algorithm based on adaptive multi-granularity information particle construction, through the whole process architecture of " data preprocessing→adaptive information particle generation→granularity feature coding→transform feature extraction→model optimization", realize the robust feature extraction of coal mine fire time sequence data.The application improves the accuracy and robustness of coal mine fire early warning significantly through whole process architecture design and technical innovation, adaptive information particle generation module is combined with local characteristic analysis and dynamic particle size adjustment strategy, generates fine-grained information particle in the area of intense data fluctuation to accurately capture early signs such as gas concentration surge, temperature anomaly, etc., while using coarse-grained summary in stable area to effectively suppress random noise, with multi-attribute feature coding, form high-dimensional discriminative feature input, make feature extraction robustness and discriminability improve more than 30% compared with traditional fixed granularity method, F1 score increases by 15%~20%.

Description

Technical Field

[0001] This invention belongs to the field of coal mine safety monitoring and early warning technology, and relates to a method for extracting the temporal features of coal mine fires by integrating adaptive multi-granularity information particles and Transformer. Background Technology

[0002] Current coal mine fire early warning technologies mainly rely on the following methods, but all of them have obvious limitations: Traditional early warning methods primarily rely on threshold alarms or simple statistical analysis based on sensor data (such as gas concentration, carbon monoxide concentration, temperature, and smoke concentration). However, the underground environment in coal mines is complex and variable, and sensor data commonly suffers from problems such as multi-source heterogeneity, noise interference, data gaps, and outliers. The diverse types of monitoring equipment in coal mines, with their varied data formats, make data integration difficult. Furthermore, the occurrence and development of fires is a dynamic evolutionary process, and its early signs may manifest as weak and complex temporal patterns, which traditional fixed threshold methods struggle to capture.

[0003] Granular computing, as an effective method for handling uncertain, imprecise, and incomplete information, has demonstrated unique advantages in data analysis and pattern recognition. By granulating raw data into information particles at different levels of abstraction, granular computing can effectively reduce noise, decrease data dimensionality, and reveal the intrinsic structure of data from both macroscopic and microscopic perspectives. However, the application of existing granular computing methods in coal mine safety monitoring still has limitations: they mostly employ fixed granularity partitioning strategies, lacking the ability to adapt to the intrinsic structure and changing patterns of the data. This fixed granularity partitioning method is like mechanically measuring text with a ruler, completely ignoring the semantic relevance of the content. It fails to capture details when data changes drastically, and introduces redundant information when data is stable, making it difficult to flexibly cope with the complexity of the coal mine environment.

[0004] In recent years, deep learning technology, especially the Transformer model, has made significant progress in processing time-series data. With its core self-attention mechanism, the Transformer can effectively capture long-term dependencies and complex nonlinear patterns in data, overcoming the gradient vanishing / exploding problem inherent in traditional recurrent neural networks when processing long sequences. However, the Transformer model still has significant shortcomings in processing time-series data from coal mine fires: First, the input representation is limited; directly receiving raw numerical values ​​or simple features as input makes it difficult to cope with the high noise characteristics of industrial data, potentially leading to the loss of crucial information. Second, traditional positional encoding methods may suffer from ambiguous positional information in long sequence processing, failing to accurately distinguish information from different locations. Furthermore, the attention mechanism may exhibit "attention bias," ignoring important information when processing high-dimensional data, resulting in insufficient robustness and discriminative power in feature extraction. Currently, the information abstraction capabilities of granular computing and the time-series modeling capabilities of the Transformer have not been effectively integrated. While multi-granularity cognitive computing has been applied in coal mine safety management, it has not been combined with the powerful temporal modeling capabilities of Transformers. Existing deep learning models, such as the CNN-Transformer hybrid model, can process temporal data but do not leverage the advantages of granular computing in information abstraction, noise reduction, and multi-granularity representation. Therefore, how to deeply integrate the adaptive information abstraction capabilities of granular computing with the powerful temporal modeling capabilities of Transformers to more effectively extract robust and discriminative early warning features from coal mine fire time-series data is a key problem urgently needing to be solved in the field of coal mine fire early warning. Summary of the Invention

[0005] The purpose of this invention is to provide a coal mine fire Transformer temporal feature extraction algorithm based on adaptive multi-granularity information particles, which solves the problems of poor robustness, insufficient ability to capture weak early signals, and poor accuracy and timeliness of early warning in the existing technology for coal mine fire temporal feature extraction.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for extracting temporal features of coal mine fires by fusing adaptive multi-granularity information particles and Transformer, comprising: S1 Data Preprocessing Module: Inputs raw sensor time-series data and outputs noise-reduced, completed, and standardized time-series data; S2 Adaptive Information Granule Generation Module: Input preprocessed standardized time series data, output adaptive information granule sequences with different granule sizes; S3 Granularity Feature Encoding Module: Inputs an adaptive information granular sequence and outputs a granular feature vector sequence containing statistical and topological attributes; S4 Transformer Feature Extraction Network: Input a sequence of granular feature vectors, output a high-order feature representation that fuses temporal dependencies; S5 Model Training and Optimization: Input high-order feature representations and true labels, output optimized model parameters and fire risk probability.

[0007] Optionally, the S1 data preprocessing module specifically includes: Step 1: Median filtering for noise reduction: ; The output value after median filtering at time point t; with t as the center, the window size is... The original data points; Median{} retrieves the median of the data within the window; Step 2: Fill in missing values ​​using cubic spline interpolation: ; The interpolation results are for the missing time points. The coefficients of the cubic spline polynomial are determined by solving a system of linear equations. The known timestamps near the missing point; Step 3, Z-score standardization: ; Preprocessed standardized data ; The mean of the training set data, The standard deviation of the training set data.

[0008] Optionally, the S2 adaptive information granule generation module specifically includes: Step 1: Initial division of the sliding window Input preprocessed standardized data The time series data is divided into equally spaced sliding windows according to the set initial window size, and the output is the initial window sequence. ; Step 2: Local Characteristic Analysis Input initial window sequence For each window, three local characteristic indicators—variance, information entropy, and data density—are calculated. The specific formulas are as follows: ; window Variance of internal data Initial window size, window The mean of the data within the dataset; Information entropy formula: ; Information entropy of the data within the window The probability distribution of the d-th dimension feature; Data density formula: ; The density of data within the window, The k-nearest neighbor index set at time point t; Euclidean distance; Both represent the standardized monitoring feature vectors; t is the current time index. This is the k nearest neighbor time index of the current sample within the window; Step 3: Conditional Judgment and Branching Input a set of local characteristic indicators, and the system will make conditional judgments based on these indicators. (1) If variance > threshold or information entropy > threshold: it indicates that the data fluctuates drastically, triggering fine-grained branching; (2) If variance < threshold and data density > threshold: it indicates that the data is stable and concentrated, triggering coarse-grained branching; (3) If the data density is less than the threshold, it indicates that the data is sparse and triggers independent granular branches; Step 4: Adaptive Particle Size Adjustment Strategy Input the local characteristic index set and the judgment result from step 3. Based on the judgment result, use the DBSCAN parameter dynamic adjustment formula to adjust the neighborhood radius. Make adjustments; reduce fine-grained branches Subdivide the window; increase the size of the coarse-grained branches. To merge data points; The formula for dynamically adjusting DBSCAN parameters is: ; The neighborhood radius of the DBSCAN algorithm. The density of data within the window, Density threshold and The minimum and maximum values ​​of the neighborhood radius.

[0009] Optionally, the S3 granularity feature encoding module includes: Step 1: Statistical Attribute Coding ; The mean of data within an information granule. Standard deviation of data within information granules Skewness of data within information granules Kurtosis of data within information granules and The maximum and minimum values ​​of data within an information granule; Step 2: Topology attribute encoding ; For information particles Time span; For information particles Density of internal data; A set of granularity indexes adjacent to the information granules; For information particles and The degree of overlap; For information particles and Adjacency relationship.

[0010] Optionally, the S4 Transformer temporal feature extraction network specifically includes: Input a sequence of granular feature vectors, output the probability of fire risk; Step 1: Input Embedding Layer ; Embedding vector of information particles; For learnable embedding matrices, For the original feature dimension, For Transformer model dimensions; Bias term; Step 2: Improved position coding: ; Encoding for absolute time and location; The time span of the information particles; and The mean and standard deviation of all information particles over the time span; Linear() linear transformation function; Step 3: Multi-head self-attention mechanism ; ; Query, key-value matrix, embedded from input Obtained by linear transformation; Dimensions of each attention head; The output of the i-th attention head, i=1, 2...h; Output the linear transformation matrix; This is the scaling factor; Step 4: Feedforward Neural Network ; , Weight matrix; , Bias term; The ReLU activation function; H represents the hidden state matrix of the input feedforward neural network; Step 5: Residual Connectivity and Layer Normalization LayerNormMultiHead; ; No. +1 layer output; LayerNorm() layer normalization function.

[0011] Optionally, the S5 model training and optimization specifically includes: Step 1: Dataset Partitioning and Preprocessing Dataset partitioning: Training set: 60% historical data; Validation set: 20% historical data; Test set: 20% historical data; Step 2: Loss Function Design ; Positive class weights are dynamically adjusted based on class distribution. Weight decay coefficient; The L2 norm of all learnable parameters of the model; N indicates that N is the sample number variable; Represents the actual label symbol; Indicates the symbol for the predicted probability; Step 3: Optimizer Selection and Parameter Settings Optimizer: AdamW ; Initial learning rate, and First and second moment estimation of the gradient; It is a small constant; This is the value of the current parameter at time step t; The updated parameters; Step 4: Regularization and Overfit Prevention Strategies Dropout: Add a Dropout layer after the FFN layer of the Transformer, randomly zeroing 30% of the neurons; Tag smoothing: Transform real tags Replace with ,in , ; Step 5: Training Process and Early Stop Mechanism Training process: Initialize model parameters; each batch of input data passes through an embedding layer, positional encoding, and a Transformer encoder; calculate the loss function and backpropagate to update model parameters; evaluate performance on the validation set every 10 epochs; early stopping mechanism: if the F1 score on the validation set does not improve for 5 consecutive epochs, stop training and retain the model with the highest F1 score.

[0012] The advantages of this invention are: This invention, through end-to-end architectural design and technological innovation, significantly improves the accuracy and robustness of early warning systems for coal mine fires. Its core beneficial effects are reflected in the following aspects: First, the adaptive information granule generation module combines local characteristic analysis (such as variance, information entropy, and data density) with dynamic granularity adjustment strategies (such as DBSCAN parameter density-sensitive adjustment). In areas of drastic data fluctuation, it generates fine-grained information granules to accurately capture early signs such as sudden increases in gas concentration and abnormal temperatures. Simultaneously, in stable areas, it employs coarse-grained generalization to effectively suppress random noise. Combined with multi-attribute feature encoding (fusion of statistical, topological, and semantic attributes), it forms high-dimensional discriminative feature inputs, improving feature extraction robustness and discriminative power by more than 30% compared to traditional fixed-granularity methods, and increasing the F1 score by 15%–20%. Second, the Transformer temporal feature extraction network, through a multi-head self-attention mechanism and improved positional encoding (fusion of absolute time and granularity information), achieves cross-granularity temporal dependency modeling and granular-temporal joint perception, significantly enhancing the early "slow-accelerate" warning of fires. The model's ability to capture complex patterns and weak signals (such as a slow increase in CO concentration) is 25% better than that of unidirectional models like LSTM, which rely on long-term capture capabilities, and it can identify early signs 10-15 minutes earlier. Furthermore, the model training employs weighted cross-entropy loss (positive class weights penalize misjudgments) and L2 regularization, combined with adversarial training, label smoothing, and Dropout strategies. This effectively addresses challenges such as high noise levels and class imbalance in coal mine data (fire samples account for <5%), reducing the false negative rate (MAR) to below 3% and the false positive rate (FAR) to below 2%, resulting in a combined false positive rate reduction of over 80% compared to traditional thresholding methods. Finally, the end-to-end optimized architecture and lightweight design (granular feature compression + efficient self-attention mechanism) improve inference speed by over 30% while maintaining accuracy (single batch processing time <0.1 seconds). The output high-order feature representation can be directly used for subsequent tasks such as fire classification and risk assessment. Combined with attention weight visualization to enhance interpretability, this provides key technical support for coal mine safety production with high accuracy, low latency, and strong robustness. Attached Figure Description

[0013] The accompanying drawings are provided to further illustrate the present disclosure and form part of the specification. They are used together with the following detailed description to explain the present disclosure, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the overall architecture of the algorithm of this invention; it shows the entire process from raw sensor data input to fire risk probability output, including data preprocessing (median filtering, interpolation, standardization), adaptive information granularity generation (sliding window, dynamic granularity adjustment), granular feature encoding (statistical / topological / semantic attribute fusion), Transformer temporal feature extraction (multi-head self-attention, position encoding) and model optimization (loss function, regularization) modules; Figure 2 The flowchart shows the adaptive information granulation generation module; it details the core steps of adaptive granulation, including sliding window partitioning, local characteristic analysis (variance, information entropy, data density calculation), dynamic granularity adjustment strategy (threshold-based window subdivision, DBSCAN parameter density sensitive adjustment), and granularity relationship encoding (inclusion, adjacency, and overlap relationships). Figure 3 This diagram shows the internal structure of the Transformer temporal feature extraction network. It also illustrates the multi-layer structure of the Transformer encoder, including the connection relationships and data flow of key components such as the input embedding layer, improved positional encoding (absolute time + granularity encoding), multi-head self-attention mechanism (8 parallel attention heads), feedforward neural network (FFN), residual connections, and layer normalization. Figure 4 The graphs show the comparison between the model training and optimization processes; they also compare the performance curves of the proposed method and the traditional LSTM / GRU model on the test set, including the trends of F1 score, false alarm rate (FAR), and false negative rate (MAR) with the number of training rounds, as well as the convergence speed of the weighted cross-entropy loss function, to verify the optimization effect of the proposed method in class imbalance scenarios. Detailed Implementation

[0014] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0015] The Transformer temporal feature extraction algorithm for coal mine fires based on adaptive multi-granularity information particles of this invention has the following overall data flow process: Overall input: raw time-series data from multiple coal mine sensors (such as gas concentration, temperature, CO concentration, etc.). Overall output: predicted risk probability value of coal mine fire.

[0016] Combination Figure 1-3 The Transformer temporal feature extraction algorithm of this invention, based on adaptive multi-granularity information granules, achieves robust feature extraction from coal mine fire time-series data through a full-process architecture of "data preprocessing → adaptive information granule generation → granular feature encoding → Transformer feature extraction → model optimization". The technical details, formula derivations, and inputs / outputs of each step are as follows: The data preprocessing module takes raw sensor time-series data (including noise and missing values) as input and outputs noise-reduced, completed, and standardized time-series data as output. The adaptive information particle generation module takes preprocessed standardized time series data as input and outputs adaptive information particle sequences with different particle sizes. The granular feature encoding module takes an adaptive information granular sequence as input and outputs a granular feature vector sequence containing statistical and topological attributes. The Transformer feature extraction network takes a sequence of granular feature vectors as input and outputs a high-order feature representation that fuses temporal dependencies. Model training and optimization: Input high-order feature representations and real labels, output optimized model parameters and fire risk probability.

[0017] Specifically, it includes: S1 Data Preprocessing Module Step 1: Median filtering for noise reduction ; The output value after median filtering at time point t; with t as the center, the window size is... The original data points (e.g.) (This represents a window with 3 data points). `Median{}` takes the median of the data within the window. Coal mine sensor data is susceptible to transient noise interference (such as electromagnetic pulses). Median filtering effectively suppresses outliers while preserving edge information by selecting the median of the data within the window. Window size It needs to be adjusted according to the data sampling frequency (e.g., when sampling every 1 minute). (Corresponding to a 5-minute window).

[0018] Step 2: Fill in missing values ​​using cubic spline interpolation formula: ; The interpolation results are for the missing time points. The coefficients of the cubic spline polynomial are determined by solving a system of linear equations. For known timestamps near the missing point (e.g.) (Or t+1). Cubic spline interpolation fits the data using piecewise cubic polynomials, ensuring the interpolation function is continuous and second-differentiable at the nodes, making it suitable for nonlinear sensor data (such as nonlinear changes in gas concentration over time). Parameters The nearest valid data point should be selected to ensure interpolation accuracy.

[0019] Step 3: Z-score standardization formula: ; Preprocessed standardized data ; The mean of the training set data. This represents the standard deviation of the training set data. It eliminates dimensional differences between different sensors (e.g., gas concentration is expressed as %, temperature as °C), ensuring the data follows a standard normal distribution, which facilitates subsequent model training and convergence. and Calculations should be based on historical data to avoid leaking future information.

[0020] S2 Adaptive Information Granule Generation Module Step 1: Initial division of the sliding window (addition) Input: Preprocessed standardized data .

[0021] Operation procedure: Divide the time series data into equally spaced sliding windows according to the set initial window size (e.g., 10 minutes), and output the initial window sequence. .

[0022] Step 2: Local Characteristic Analysis Input: Initial window sequence .

[0023] Operation process: Calculate three local characteristic indicators—variance, information entropy, and data density—for the data within each window. The specific formulas are as follows: Variance formula: ; window Variance of internal data Initial window size (e.g., 10 minutes), window The mean of the data. Variance measures the volatility of the data. High variance indicates drastic changes in the data (such as a sudden increase in gas concentration), requiring fine-grained noise reduction; low variance indicates stable data, which can be reduced using coarse-grained noise reduction.

[0024] ; The lower bound for summation represents the starting time index of the i-th sliding window. Here, t is the time index in the time series data, i is the window number, S is the sliding step size, and the initial window size is represented by W plus the indices of init in the formula above. Therefore, the i-th window covers the standardized samples from the starting index to the starting index plus the window length minus one. This definition explains the origin of the upper and lower bounds for summation in the variance formula.

[0025] • Information entropy formula: ; Information entropy of the data within the window The probability distribution of the d-th dimension feature (such as gas concentration). Information entropy quantifies data uncertainty. High entropy indicates high data complexity (such as the superposition of multiple noise sources), requiring fine-grained analysis; low entropy indicates strong data regularity, allowing for coarse-grained generalization.

[0026] Data density formula: ; The density of data within the window, The k-nearest neighbor index set at time point t (e.g., k=5); Euclidean distance. Both represent the standardized monitoring feature vectors; t is the current time index. This is the k-nearest neighbor time index of the current sample within the window; the norm term represents the Euclidean distance between the two neighbors, used to calculate the data density within the window. Data density reflects the local data density. Low-density regions may be outliers (such as sensor malfunctions) and need to be treated as independent granular points; high-density regions represent concentrated data and can be summarized using coarse-grained methods.

[0027] Output: A set of local characteristic indicators including variance, information entropy, and density.

[0028] Step 3: Conditional Judgment and Branch Path (corresponding to the logic in Figure 2) Input: Local characteristic index set.

[0029] Judgment Logic: As shown in Figure 2, the system makes conditional judgments based on local characteristic indicators: (1) If variance > threshold or information entropy > threshold: it indicates that the data fluctuates drastically (such as gas mutation), triggering fine-grained branching.

[0030] (2) If variance < threshold and data density > threshold: it indicates that the data is stable and concentrated, triggering coarse-grained branching.

[0031] (3) If the data density is less than the threshold, it indicates that the data is sparse (possibly an outlier) and triggers independent granular branches.

[0032] Output: The result of determining the granularity adjustment direction for each window.

[0033] Step 4: Adaptive Particle Size Adjustment Strategy Input: Local characteristic index set and the judgment result of step 3.

[0034] Operation process: Based on the judgment result, the neighborhood radius is dynamically adjusted using the DBSCAN parameter adjustment formula. Adjustments were made. The fine-grained branches were reduced. Subdivide the window; increase the size of the coarse-grained branches. To merge data points.

[0035] DBSCAN parameter dynamic adjustment formula ; The neighborhood radius of the DBSCAN algorithm. The density of data within the window, Density threshold (e.g.) , and Minimum and maximum values ​​of neighborhood radius (e.g.) In low-density areas Decrease This generates smaller granularities to capture sparse but crucial information (such as temperature anomalies); and increases granularity in high-density regions. Neighboring points are merged to reduce noise. Output: The final generated sequence of information granularities with an adaptive size is passed to the granular feature encoding module.

[0036] S3 Granularity Feature Encoding Module Step 1: Statistical Attribute Coding ; The mean of data within an information granule. Standard deviation of data within information granules Skewness of data within information granules (reflecting distribution asymmetry), : Kurtosis of data within an information granularity (reflecting the sharpness of the distribution), and The maximum and minimum values ​​of data within an information granule. Statistical attributes comprehensively describe the numerical characteristics of the information granule. For example, high skewness may indicate that the data is shifted to the right (such as an abnormal increase in gas concentration), while kurtosis may indicate that the data is concentrated in a certain area (such as temperature stabilizing within a certain range).

[0037] Step 2: Topology attribute encoding ; Information Particle The time span (e.g., 10 minutes). Information Particle Density of internal data. The set of granularity indexes adjacent to the information granules. Information Particle and The degree of overlap (the proportion of overlap in the feature space). Information Particle and The adjacency relationship is determined by the temporal proximity (1 for temporally adjacent, 0 otherwise). Topological properties provide the spatiotemporal relationships of information granularities. For example, long time spans indicate stable data, while short time spans indicate abrupt changes in data; adjacency overlap indicates the correlation between granularities, providing prior information for the Transformer attention mechanism.

[0038] S4 Transformer Temporal Feature Extraction Network Input: A sequence of granular feature vectors. Output: Fire risk probability.

[0039] Step 1: Input Embedding Layer ; Embedding vector of information particles. : A learnable embedding matrix, where For the original feature dimension, For Transformer model dimensions (e.g., 512). Bias term. Maps granular feature vectors to a higher-dimensional space, enhancing the model's expressive power. For example, if (statistical attributes 6-dimensional + topological attributes 4-dimensional + semantic attributes 1-dimensional), then it is extended to dimensionality through linear transformation.

[0040] Step 2: Improved Position Coding ; Absolute time position encoding (sine / cosine function generation). The time span of information particles. and Mean and standard deviation of all information granular time spans (calculated based on historical data). Linear() linear transformation function (e.g., fully connected layer). Absolute time encoding: preserves the position information of the granularity in the global time series (e.g., the 10th granularity corresponds to the 100th minute). Granularity size encoding: linearly maps the time span after standardization, allowing the model to distinguish different granularity levels (e.g., fine granularity corresponds to short time spans, coarse granularity corresponds to long time spans).

[0041] Step 3: Multi-head self-attention mechanism ; ; Query, key-value matrix, embedded from input Obtained by linear transformation. Dimensions of each attention head (e.g.) ). The output of the i-th attention head, i=1, 2...h. Outputs a linear transformation matrix. Multi-head mechanism: Eight parallel attention heads capture inter-granular relationships from different subspaces (e.g., head 1 focuses on gas particle size changes, head 2 focuses on temperature particle size correlations). Scaling factor: To prevent the gradient from vanishing due to an excessively large inner product.

[0042] Step 4: Feedforward Neural Network ; , Weight matrix. , Bias term. The ReLU activation function enhances the expressive power of a model through nonlinear transformations. By introducing sparsity, it improves computational efficiency. For example, it maps 512-dimensional features to 2048 dimensions and then back to 512 dimensions, thus enhancing feature diversity. The hidden state matrix represents the input to the feedforward neural network, which is the intermediate representation of the information particle feature sequence after input embedding, position encoding, multi-head self-attention and residual normalization. Its dimensional relationship is shown in the supplementary formula, where m is the number of information particles and the model dimension symbol represents the Transformer model dimension. ; Step 5: Residual Connectivity and Layer Normalization LayerNormMultiHead; ; The output of layer (l+1). LayerNorm() is the layer normalization function. Residual connections: alleviate the vanishing gradient problem in deep networks and accelerate training convergence. Layer normalization: stabilizes the input distribution of each layer and improves model robustness.

[0043] S5 Model Training and Optimization Objective: To improve the accuracy and robustness of early fire symptom identification by optimizing model parameters through end-to-end training, while addressing challenges such as high noise and class imbalance in coal mine data.

[0044] Step 1: Dataset Partitioning and Preprocessing Dataset partitioning: Training set: 60% historical data (including normal and early fire symptom samples). Validation set: 20% historical data (for hyperparameter tuning). Test set: 20% historical data (for final performance evaluation). Data augmentation: Adversarial training: Adding random noise (such as Gaussian noise) to the input features to improve the model's robustness to noise. Mixed sampling: Using SMOTE oversampling for the minority class (fire samples) and random undersampling for the majority class (normal samples) to alleviate class imbalance.

[0045] Step 2: Loss Function Design ; Positive class weights (e.g.) ), dynamically adjusted based on category distribution. Weight decay coefficient (e.g.) ), which controls the strength of L2 regularization. The L2 norm of all learnable parameters of the model. Weighted cross-entropy: increased by increasing the weights of the positive class. This penalizes the model for misclassifying fire samples (reducing the false negative rate). L2 regularization: prevents the model from overfitting and improves generalization ability. N indicates that N is a variable representing the number of samples; Represents the actual label symbol; Indicates the symbol for the predicted probability; Step 3: Optimizer Selection and Parameter Settings Optimizer: AdamW (Adam with Weight Decay) ; Initial learning rate (e.g.) =0.001), and First and second moment estimation of gradient (with bias correction). Small constants (such as) =10 -8 This prevents division by zero errors. This represents the value of the current parameter (weight or bias) at time step t. These are the updated parameters.

[0046] AdamW: Combining Adam's momentum estimation with L2 regularization, it effectively balances training speed and generalization ability. Learning rate decay: It adopts a cosine annealing strategy, which decays the learning rate to 90% of the original value every 10 epochs.

[0047] Step 4: Regularization and Overfit Prevention Strategies 1. Dropout: Add a Dropout layer after the FFN layer of the Transformer, randomly zeroing out 30% of the neurons (Dropout rate = 0.3). This prevents feature dependencies from becoming overly concentrated on specific neurons, improving model robustness.

[0048] 2. Tag smoothing: Smooths the actual tags. Replace with ,in (Binary classification) This helps alleviate model overfitting and increases tolerance to noise labels.

[0049] Step 5: Training Process and Early Stop Mechanism Training process: 1. Initialize model parameters (using Xavier initialization). 2. Each batch of input data passes through an embedding layer, positional encoding, and a Transformer encoder. 3. Calculate the loss function and backpropagate to update model parameters. 4. Evaluate performance (F1 score) on the validation set every 10 epochs. Early stopping mechanism: If the F1 score on the validation set does not improve for 5 consecutive epochs, stop training and retain the model with the highest F1 score. This prevents overfitting and ensures the model achieves optimal performance on the validation set.

[0050] Step 6: Model Evaluation Metrics Key metrics: F1 score: Balances precision and recall, suitable for imbalanced class scenarios. AUC-ROC: Measures the model's generalization ability to different thresholds. Auxiliary metrics: False Alarm Rate (FAR): The proportion of normal samples misclassified as fires. False Negative Rate (MAR): The proportion of fire samples missed.

[0051] This invention, through end-to-end architecture design and technological innovation, significantly improves the accuracy and robustness of early warning systems for coal mine fires. Its core beneficial effects are reflected in the following aspects: First, the adaptive information granule generation module combines local characteristic analysis (such as variance, information entropy, and data density) with dynamic granularity adjustment strategies (such as density-sensitive adjustment of DBSCAN parameters). In areas of drastic data fluctuation, it generates fine-grained information granules to accurately capture early signs such as sudden increases in gas concentration and abnormal temperatures. Simultaneously, in stable areas, it employs coarse-grained generalization to effectively suppress random noise. Combined with multi-attribute feature encoding (fusion of statistical, topological, and semantic attributes), it forms high-dimensional discriminative feature inputs, improving feature extraction robustness and discriminative power by more than 30% compared to traditional fixed-granularity methods, and increasing the F1 score by 15%-20%. Second, the Transformer temporal feature extraction network, through a multi-head self-attention mechanism and improved positional encoding (fusion of absolute time and granularity information), achieves cross-granularity temporal dependency modeling and granular-temporal joint perception, significantly enhancing the early "slow-accelerate" warning of fires. The model's ability to capture complex patterns and weak signals (such as a slow increase in CO concentration) is 25% better than that of unidirectional models like LSTM, which rely on long-term capture capabilities, and it can identify early signs 10-15 minutes earlier. Furthermore, the model training employs weighted cross-entropy loss (with positive class weights penalizing misjudgments) and L2 regularization, combined with adversarial training, label smoothing, and Dropout strategies. This effectively addresses challenges such as high noise levels and class imbalance in coal mine data (fire samples account for <5%), reducing the false negative rate (MAR) to below 3% and the false positive rate (FAR) to below 2%, a reduction of over 80% in the overall false positive rate compared to traditional thresholding methods. Finally, the end-to-end optimized architecture and lightweight design (granular feature compression + efficient self-attention mechanism) improve inference speed by over 30% while maintaining accuracy (single batch processing time <0.1 seconds). The output high-order feature representation can be directly used for subsequent tasks such as fire classification and risk assessment. Combined with attention weight visualization to enhance interpretability, this provides key technical support for high-precision, low-latency, and robust coal mine safety production.

[0052] The above-mentioned technical effects have been verified on a real historical monitoring dataset from a large coal mine. This dataset contains minute-by-minute sampling data from multiple sensors (gas, CO, temperature) for six consecutive months, totaling approximately 260,000 records, including three abnormal fluctuation samples from the early stages of real fires. We conducted comparative experiments on the adaptive multi-granularity Transformer model proposed in this invention with traditional fixed-granularity LSTM and CNN-Transformer models. The specific comparison results are shown in the table below:

[0053] Experimental results show that, under the same test set, the model of this invention can identify weak early fire features 12 minutes earlier, with an F1 score of 0.92 (an improvement of about 18% compared to LSTM). At the same time, during the stable period before the fire occurs, the false alarm rate is only 1.5%, which verifies its excellent performance in complex coal mine environments.

[0054] Figure 4 The results were obtained through comparative training experiments on a historical monitoring dataset of a large coal mine. Under the same data partitioning, preprocessing method, training rounds, and evaluation criteria, the proposed method, LSTM, GRU, and CNN-Transformer models were trained respectively, and the F1 score, false alarm rate (FAR), false negative rate (MAR), and weighted cross-entropy loss value were recorded in each training round. Figure 4 The results show that the F1 score of the method of the present invention improves faster and reaches the highest value, while the FAR, MAR and loss values ​​decrease faster and reach the lowest value. This indicates that the method has higher accuracy, lower false alarm and false negative risk and better training convergence performance in early fire feature recognition.

[0055] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A method for extracting temporal features of coal mine fires by fusing adaptive multi-granularity information particles and Transformer, characterized in that, include: S1 Data Preprocessing Module: Inputs raw sensor time-series data and outputs noise-reduced, completed, and standardized time-series data; S2 Adaptive Information Granule Generation Module: Input preprocessed standardized time series data, output adaptive information granule sequences with different granule sizes; S3 Granularity Feature Encoding Module: Inputs an adaptive information granular sequence and outputs a granular feature vector sequence containing statistical and topological attributes; S4 Transformer Feature Extraction Network: Input a sequence of granular feature vectors, output a high-order feature representation that fuses temporal dependencies; S5 Model Training and Optimization: Input high-order feature representations and true labels, output optimized model parameters and fire risk probability.

2. The method for extracting temporal features of coal mine fires by fusing adaptive multi-granularity information particles and Transformer as described in claim 1, characterized in that, The S1 data preprocessing module specifically includes: Step 1: Median filtering for noise reduction: ； The output value after median filtering at time point t; with t as the center, the window size is... The original data points; Median{} retrieves the median of the data within the window; Step 2: Fill in missing values ​​using cubic spline interpolation: ； The interpolation results are for the missing time points. The coefficients of the cubic spline polynomial are determined by solving a system of linear equations. The known timestamps near the missing point; Step 3, Z-score standardization: ； Preprocessed standardized data ; The mean of the training set data, The standard deviation of the training set data.

3. The method for extracting temporal features of coal mine fires by fusing adaptive multi-granularity information particles and Transformer according to claim 1 or 2, characterized in that, The S2 adaptive information granule generation module specifically includes: Step 1: Initial division of the sliding window Input preprocessed standardized data The time series data is divided into equally spaced sliding windows according to the set initial window size, and the output is the initial window sequence. ; Step 2: Local Characteristic Analysis Input initial window sequence For each window, three local characteristic indicators—variance, information entropy, and data density—are calculated. The specific formulas are as follows: ； window Variance of internal data Initial window size, window The mean of the data within the dataset; Information entropy formula: ； Information entropy of the data within the window The probability distribution of the d-th dimension feature; Data density formula: ； The density of data within the window, The k-nearest neighbor index set at time point t; Euclidean distance; Both represent the standardized monitoring feature vectors; t is the current time index. This is the k nearest neighbor time index of the current sample within the window; Step 3: Conditional Judgment and Branching Input a set of local characteristic indicators, and the system will make conditional judgments based on these indicators. (1) If variance > threshold or information entropy > threshold: it indicates that the data fluctuates drastically, triggering fine-grained branching; (2) If variance < threshold and data density > threshold: it indicates that the data is stable and concentrated, triggering coarse-grained branching; (3) If the data density is less than the threshold, it indicates that the data is sparse and triggers independent granular branches; Step 4: Adaptive Particle Size Adjustment Strategy Input the local characteristic index set and the judgment result from step 3. Based on the judgment result, use the DBSCAN parameter dynamic adjustment formula to adjust the neighborhood radius. Make adjustments; reduce fine-grained branches Subdivide the window; increase the size of the coarse-grained branches. To merge data points; The formula for dynamically adjusting DBSCAN parameters is: ； The neighborhood radius of the DBSCAN algorithm. The density of data within the window, Density threshold and The minimum and maximum values ​​of the neighborhood radius.

4. The method for extracting temporal features of coal mine fires by fusing adaptive multi-granularity information particles and Transformer according to claim 1 or 2, characterized in that, The S3 granularity feature encoding module includes: Step 1: Statistical Attribute Coding ； The mean of data within an information granule. Standard deviation of data within information granules Skewness of data within information granules Kurtosis of data within information granules and The maximum and minimum values ​​of data within an information granule; Step 2: Topology attribute encoding ； For information particles Time span; For information particles Density of internal data; A set of granularity indexes adjacent to the information granules; For information particles and The degree of overlap; For information particles and Adjacency relationship.

5. The method for extracting temporal features of coal mine fires by fusing adaptive multi-granularity information particles and Transformer according to claim 1 or 2, characterized in that, The S4 Transformer temporal feature extraction network specifically includes: Input a sequence of granular feature vectors, output the probability of fire risk; Step 1: Input Embedding Layer ； Embedding vector of information particles; For learnable embedding matrices, For the original feature dimension, For Transformer model dimensions; Bias term; Step 2: Improved position coding: ； Encoding for absolute time and location; The time span of the information particles; and The mean and standard deviation of all information particles over the time span; Linear() linear transformation function; Step 3: Multi-head self-attention mechanism ； ； Query, key-value matrix, embedded from input Obtained by linear transformation; Dimensions of each attention head; The output of the i-th attention head, i=1, 2...h; Output the linear transformation matrix; This is the scaling factor; Step 4: Feedforward Neural Network ； , Weight matrix; , Bias term; The ReLU activation function; H represents the hidden state matrix of the input feedforward neural network; Step 5: Residual Connectivity and Layer Normalization LayerNormMultiHead; ； No. +1 layer output; LayerNorm() layer normalization function.

6. The method for extracting temporal features of coal mine fires by fusing adaptive multi-granularity information particles and Transformer according to claim 1 or 2, characterized in that, The S5 model training and optimization specifically includes: Step 1: Dataset Partitioning and Preprocessing Dataset partitioning: Training set: 60% historical data; Validation set: 20% historical data; Test set: 20% historical data; Step 2: Loss Function Design ； Positive class weights are dynamically adjusted based on class distribution. Weight decay coefficient; The L2 norm of all learnable parameters of the model; N indicates that N is the sample number variable; Represents the actual label symbol; Indicates the symbol for the predicted probability; Step 3: Optimizer Selection and Parameter Settings Optimizer: AdamW ； Initial learning rate, and First and second moment estimation of the gradient; It is a small constant; This is the value of the current parameter at time step t; The updated parameters; Step 4: Regularization and Overfit Prevention Strategies Dropout: Add a Dropout layer after the FFN layer of the Transformer, randomly zeroing 30% of the neurons; Tag smoothing: Transform real tags Replace with ,in , ; Step 5: Training Process and Early Stop Mechanism Training process: Initialize model parameters; each batch of input data passes through an embedding layer, positional encoding, and a Transformer encoder; calculate the loss function and backpropagate to update model parameters; evaluate performance on the validation set every 10 epochs; early stopping mechanism: if the F1 score on the validation set does not improve for 5 consecutive epochs, stop training and retain the model with the highest F1 score.

Images