Gas concentration prediction method based on multi-scale feature fusion of irregular sampling data

Abstract

The present application relates to a kind of gas concentration prediction method based on the multiscale feature fusion of irregular sampling data, belong to coal mine gas concentration prediction technical field.The method includes: S1: data preparation: according to the shift time of coal mine, carry out data sequence division;S2: continuous time coding and feature fusion: using neural differential to the division sequence potential space mapping;Division multi-time window, carry out multiscale data feature extraction and fusion;S3: shift adaptive design: using hybrid expert structure and gate network, carry out adaptive processing to the change of coal mine shift, expert layer output weighted sum feature;S4: failure fault-tolerant design: carry out failure fault-tolerant processing, carry out compensation to the data missing problem caused by common sensor anomaly, obtain reconstruction feature;S5: the expert layer output feature and reconstruction feature are spliced, input prediction network, predict the gas concentration of next time.The present application can significantly improve the accuracy of toxic and harmful gas.

Description

Technical Field

[0001] This invention belongs to the field of coal mine gas concentration prediction technology, and relates to a gas concentration prediction method based on multi-scale feature fusion of irregular sampling data. Background Technology

[0002] In the safety monitoring of coal mining faces, accurate prediction of the concentration of toxic and harmful gases (especially methane) is a core element in accident prevention. However, current technologies and methods in this field suffer from three major bottlenecks, making it difficult to meet the needs of intelligent and refined safety management in modern coal mines.

[0003] 1) The Challenges of Processing and Fusion of Multi-Source Heterogeneous Data: Existing technologies face significant challenges in processing multi-dimensional data from coal mining faces. The working face gathers various types of data with vastly different characteristics:

[0004] Uneven sampling frequency: The sampling rates of different sensors vary significantly. For example, the sampling period for monitoring data such as coal and rock pressure, gas concentration, and wind speed can range from seconds to minutes. Traditional methods typically employ resampling techniques (such as interpolation or downsampling) to unify all data to the same time frequency (e.g., minutes). However, this directly leads to the loss of valuable information in high-frequency data and the distortion of the temporal characteristics of low-frequency data.

[0005] Data type heterogeneity: There are fundamental differences between dynamic monitoring data (such as gas concentration and wind speed), static geological data (such as manually measured gas content and coal seam thickness), and dynamic operating condition data (such as coal mining machine position and advance speed). Existing methods often only perform simple feature stitching, lacking effective cross-domain feature interaction and deep fusion mechanisms, and cannot fully explore the inherent correlations between different data sources.

[0006] 2) Limitations of adaptive modeling for working conditions: Coal mining processes have significant periodicity, typically divided into mining shifts and maintenance shifts. The gas generation and migration mechanisms under these two shifts are drastically different.

[0007] Mining Team: The high-speed operation of large equipment such as coal mining machines and strong mechanical vibrations will intensify the desorption and outburst of gas in the coal seam, resulting in a sharp change in gas concentration.

[0008] Maintenance team: Most of the equipment is in a shutdown or low-speed maintenance state, the working surface is less disturbed, and the gas outflow is relatively gentle, but poor local ventilation may lead to gas accumulation.

[0009] However, traditional prediction models widely used in the industry, such as Autoregressive Integral Moving Average (ARIMA), Long Short-Term Memory (LSTM), and Transformer, typically employ a globally uniform modeling strategy. These models cannot adaptively adjust their internal parameters and prediction logic according to shift changes, resulting in a significant decrease in prediction accuracy or even complete failure when critical shift transitions or specific operational events (such as coal mining machine start / stop) occur.

[0010] 3) Vulnerability of Model Failure Due to Sensor Failure: The harsh environment of underground coal mines, with high temperatures, high humidity, and heavy dust, leads to a high failure rate for sensors, especially critical sensors such as gas probes. When these core sensors experience sudden failures and cannot provide effective data, existing prediction models will suffer severe prediction drift due to missing or abnormal inputs, outputting erroneous information that is of no reference value or even misleading, posing a significant threat to safe production.

[0011] In summary, the resampling strategies, globally unified modeling paradigms, and high dependence on the integrity of sensor data commonly used in existing technologies constitute the current technical bottlenecks in the field of toxic and hazardous gas prediction in coal mines. These methods cannot effectively handle non-uniform sampling data, cannot adaptively distinguish between different production conditions, and lack robustness in the event of critical sensor failures. Their prediction accuracy and reliability no longer meet the stringent requirements of the "Coal Mine Safety Regulations." Therefore, there is an urgent need to develop a novel prediction method to overcome the limitations of existing technologies and improve the level of intelligent safety production in coal mines. Summary of the Invention

[0012] In view of this, the purpose of this invention is to provide a gas concentration prediction method based on multi-scale feature fusion of irregular sampling data. This method overcomes the information loss caused by resampling due to differences in sensor sampling frequencies (from 1 second to several minutes), achieving efficient feature extraction from native irregular time-series data. It breaks through the limitations of traditional globally unified models by using shift-driven segmented modeling to capture the unique patterns of mining and maintenance shifts separately. It also solves the prediction drift caused by sudden sensor failures in the well by establishing a dynamic mask training and cross-sensor feature compensation mechanism. This invention can simultaneously solve the problems of non-uniform sampling, adaptive modeling under operating conditions, and sensor fault tolerance, significantly improving the accuracy of predictions for toxic and harmful gases.

[0013] To achieve the above objectives, the present invention provides the following technical solution:

[0014] A gas concentration prediction method based on multi-scale feature fusion of irregularly sampled data includes the following steps:

[0015] S1: Data preparation: Divide the data sequence according to the shift schedule set by the coal mine;

[0016] S2: Continuous-time encoding and feature fusion: Applying neural constant differential to the partitioned sequence for latent space mapping; dividing into multiple time windows for multi-scale data feature extraction and fusion;

[0017] S3: Shift Adaptive Design: Adopts a hybrid expert structure and gating network to adaptively handle changes in coal mine shifts, with the expert layer outputting weighted sum features;

[0018] S4: Failure-Tolerant Design: Implement failure-tolerant design to compensate for data loss caused by common sensor anomalies and obtain reconstructed features;

[0019] S5: Concentration Prediction: The weighted sum features output by the expert layer are concatenated with the reconstructed features and input into the prediction network to output the predicted gas concentration.

[0020] Furthermore, step S1 specifically includes: according to the coal mine shift (mining shift / maintenance shift) schedule, dividing the multi-source sensor data stream into mining sub-sequences and maintenance sub-sequences according to the shift, retaining the timestamps, and not resampling.

[0021] Furthermore, step S2 specifically includes: continuous time encoding and feature fusion: using a neural ordinary differential equation encoder to map multi-source measurements at irregular time points to a continuous latent space; dividing a sliding window to extract multi-scale features, concatenating features from different time scales into vectors, and fusing and normalizing them through a fully connected layer.

[0022] Furthermore, in step S2, a neural ordinary differential equation encoder is used to map multi-source measurements at irregular time points to a continuous latent space, specifically including:

[0023]

[0024] Where h(t) is the hidden state at time t, x(t) is the observed input data at time t, and f θ (·) represents the ordinary differential equation, h(t0) represents the hidden state at time t0, x(t0) represents the observation input data at time t0, and MLP(·) represents the multilayer perceptron.

[0025] Using the RK4 constant step size ODE solver, h(t) is integrated from t1 to all sampling times t. i The hidden state sequence H is obtained;

[0026]

[0027] in, To represent the N*d dimensional real number field, i.e., all real values. N is the length of the time series, and d is the dimension of the hidden state vector at each time point;

[0028] Multi-scale feature extraction is performed by dividing the window into sliding windows. For window w, at each time point t... i Take all elements falling within the interval [t] i -w,t i The hidden state subset S of ] (w) ;

[0029] S (w) ={h(t) i )∣t i ∈[t i -w,t i ]}

[0030] For S (w) Perform self-attention pooling with positional encoding:

[0031] Q = W Q h(t N ), K = W K [S (w) ], V=W V [S (w) ]

[0032]

[0033] Where Q is the vector obtained through the hidden state; K is used to perform click calculations with the query Q vector to evaluate the matching degree; V is the value vector used to obtain the final output; W Q W K W V Here, f represents the weight matrices for vectors Q, K, and V, respectively, α is the weight vector for the attention mechanism, and f (w) This is the final output vector.

[0034] Furthermore, step S3 specifically includes: designing a hybrid expert architecture, employing a two-way expert network, including a mining expert network and a maintenance expert network, wherein the mining expert network is used to handle high-frequency, drastic fluctuation scenarios, and the maintenance expert network is used to handle low-frequency, slowly changing accumulation scenarios; constructing a gating network to output routing weights; and finally adding an expert layer to output weighted sum features.

[0035] Furthermore, in step S3, each expert network is a multi-layer Transformer used to capture long-short-term dependencies and nonlinear features.

[0036] Furthermore, step S4 specifically includes: during the model training phase, randomly masking each sensor channel with the input features to simulate probe failure scenarios; adding a reconstruction sub-network branch to the loss function, so that when a channel is masked, the reconstruction sub-network uses features from other channels F... -k Estimate missing channels And weighted recovery; the reconstructed subnetwork adopts a lightweight residual structure:

[0037]

[0038] Where, x res It represents the residual.

[0039] Furthermore, in step S5, the loss function of the prediction network is:

[0040]

[0041] in, To predict network loss, To predict the future t of the network output n+1 Predicted concentration value at time y n+1 For t n+1 The actual concentration value at time x, where λ is the regularization parameter, and x (k) This represents the data collected by the k-th sensor.

[0042] Furthermore, in step S5, the prediction network uses a common network such as LSTM, GRU, or Transformer.

[0043] The beneficial effects of this invention are as follows:

[0044] 1) Efficient processing of irregularly sampled data: This scheme uses neural network constant differential equations to perform continuous-time encoding on irregularly sampled multi-source sensor data, achieving efficient feature extraction of native irregular time-series data and avoiding information loss and temporal distortion caused by traditional resampling methods. This method not only preserves high- and low-frequency dynamic information but also enhances the ability to capture abrupt changes and trends.

[0045] 2) Dynamic Adaptation to Different Operating Conditions: By using shift-driven segmented modeling, this scheme can capture the unique patterns of mining shifts (gas surge) and maintenance shifts (gas accumulation) separately, and achieve dynamic switching through a hybrid expert network and gated routing mechanism. This enables the model to adapt to changes in different operating conditions, significantly reducing false alarms and false negatives, and improving the accuracy and robustness of predictions.

[0046] 3) Sensor Fault Tolerance Mechanism: To address the high rate of sudden sensor failures in harsh downhole environments, this solution establishes a dynamic mask training and cross-sensor feature compensation mechanism. By using a channel masking strategy and a reconstructed subnetwork to compensate for critical sensor failures, the prediction drift problem caused by sensor failures is resolved, improving the system's stability and reliability.

[0047] In summary, this solution possesses significant technical advantages. By employing continuous-time encoding and multi-scale feature fusion of irregularly sampled data, shift-adaptive design, and a failure-to-tolerance mechanism, this solution effectively addresses issues such as large differences in sensor sampling frequencies, frequent changes in operating conditions, and sensor failures in underground coal mines, thereby significantly improving the accuracy and reliability of gas concentration prediction. This not only helps prevent gas disasters but also ensures the safety of miners and enhances the overall safety of coal mine production.

[0048] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0049] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0050] Figure 1 This is a flowchart of the gas concentration prediction method based on multi-scale feature fusion of irregular sampling data according to the present invention. Detailed Implementation

[0051] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0052] Please see Figure 1This invention provides a gas concentration prediction method based on multi-scale feature fusion of irregularly sampled data. The overall idea is as follows: According to the coal mine shift schedule (mining shift / maintenance shift), the sensor data stream is divided into two sub-sequences according to the shift. The original timestamp of each sensor data is retained. The irregularly sampled multi-source sensor data is continuously time-encoded using a neural network and the fusion features are extracted on a multi-scale window. Then, in the shift adaptive layer, a hybrid expert network and a gated routing mechanism are used to realize the dynamic switching between mining and maintenance conditions. Finally, in the failure tolerance layer, a channel masking strategy and a reconstructed sub-network are used to compensate for the failure of key sensors. The recovered features and expert outputs are input into the final prediction network, and the gas concentration at the next moment is accurately predicted by joint training.

[0053] This method specifically includes the following parts:

[0054] (1) Data preparation

[0055] Based on the coal mine shift schedule (mining shift / maintenance shift), the sensor data stream is divided into mining subsequences D according to the shift. mine and maintenance subsequence D maint Retain timestamps and do not resample. Assume the data from the k-th sensor is...

[0056]

[0057] Then the sensor data during the mining shift is

[0058]

[0059] The sensor data for the maintenance shift is as follows:

[0060]

[0061] (2) Continuous-time coding and feature fusion

[0062] A neural constant differential equation encoder is used to map multi-source measurements at irregular time points to a continuous latent space:

[0063]

[0064] Using the RK4 constant step size ODE solver, h(t) is integrated from t1 to all sampling times t. i This yields the hidden state sequence.

[0065]

[0066] Multi-scale feature extraction is performed by dividing the window into sliding windows, typically set to w = [30min, 60min, 480min]. For window w, at each time point t... i Take all elements falling within the interval [t] i -w,t i The hidden state subset S of ] (w) .

[0067] S (w) ={h(t) i )∣t i ∈[t i -w,t i ]}

[0068] For S (w) Perform self-attention pooling with positional encoding:

[0069] Q = W Q h(t N ), K = W K [S (w) ], V=W V [S (w) ]

[0070]

[0071] Concatenate features from different time scales into vectors:

[0072]

[0073] Fusion and normalization are performed using a fully connected layer:

[0074]

[0075] Among them, W fc Let b be the weight matrix. fc , where is the bias term, LayerNorm is the layer normalization, and ReLU is the activation function.

[0076] (3) Adaptive Class Schedule Design

[0077] Design a hybrid expert architecture, employing a two-way expert network: a mining expert network E mine Proficient in handling high-frequency, drastic fluctuation scenarios; maintenance expert network E maint Suitable for low-frequency, slowly varying accumulation scenarios. Each expert network is a multi-layer Transformer used to capture long- and short-term dependencies and nonlinear features.

[0078] Construct a gated network g and output the routing weights α;

[0079]

[0080] Where σ is the sigmoid function and MLP is a multilayer perceptron.

[0081] The final weighted sum feature E output by the expert layer is:

[0082]

[0083] (4) Failure-tolerant design

[0084] During the model training phase, the input features F are randomly masked (set to zero) for each sensor channel according to a certain probability to simulate probe failure. A reconstruction sub-network branch is added to the loss function. When a channel is masked, the reconstruction sub-network uses features F from other channels. -k Estimate missing channels And weighted recovery. The reconstructed subnetwork adopts a lightweight residual structure:

[0085]

[0086] (5) Concentration prediction

[0087] The weighted sum feature E output by the expert layer is concatenated with the reconstructed feature and input into the final prediction network P to output the future [t]. n+1, t n+2, …,t n+n ]of The prediction network here can be a general network such as LSTM, GRU, or Transformer.

[0088] The loss function is:

[0089] The method of this invention significantly improves gas prediction performance through multi-scale fusion of irregular data: it can handle the unequal time series commonly found in underground coal mine sensors, retain high and low frequency dynamic information, and enhance the ability to capture sudden changes / trends; it can dynamically adapt to the working conditions of mining shifts (gas surge) and maintenance shifts (gas accumulation), reducing false alarm and missed alarm rates; and it can solve the problem of data loss caused by common sensor failures.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A gas concentration prediction method based on multi-scale feature fusion of irregularly sampled data, characterized in that, The method specifically includes the following steps: S1: Data preparation: According to the shift schedule set by the coal mine, the data sequence is divided. Specifically, according to the coal mine shift schedule, the multi-source sensor data stream is divided into mining sub-sequence and maintenance sub-sequence according to the shift, the timestamp is retained, and no resampling is performed. S2: Continuous-time encoding and feature fusion: Applying neural constant differential to the partitioned sequence for latent space mapping; dividing the sliding window to perform multi-scale data feature extraction and fusion; S3: Shift Adaptive Design: Adopts a hybrid expert structure and gating network to adaptively handle changes in coal mine shifts, with the expert layer outputting weighted sum features; S4: Failure Tolerance Design: Failure tolerance is implemented to compensate for data loss caused by common sensor anomalies, resulting in reconstructed features. During model training, the input features are randomly masked for each sensor channel to simulate probe failure. A reconstruction sub-network branch is added to the loss function; when a channel is masked, the reconstruction sub-network uses features from other channels. Estimate missing channels And weighted recovery; the reconstructed subnetwork adopts a lightweight residual structure: in, For residuals; S5: Concentration Prediction: The weighted sum features output by the expert layer are concatenated with the reconstructed features and input into the prediction network to predict the gas concentration at the next moment.

2. The gas concentration prediction method according to claim 1, characterized in that, Step S2 specifically includes: continuous time encoding and feature fusion: using a neural frequent differential equation encoder to map multi-source measurements at irregular time points to a continuous latent space; dividing a sliding window to extract multi-scale features, concatenating features from different time scales into vectors, and fusing and normalizing them through a fully connected layer.

3. The gas concentration prediction method according to claim 2, characterized in that, In step S2, a neural constant differential equation encoder is used to map multi-source measurements at irregular time points to a continuous latent space, specifically including: in, For time points The hidden state, For time The observation input data, For God's frequent differential equations, To indicate a point in time The hidden state, For time The observation input data, It is a multilayer perceptron; Using the RK4 variable step size ODE solver, from Integrate over all sampling times The hidden state sequence is obtained. ; in, To represent the N*d-dimensional real number field, that is, all real values; The length of the time series. d The dimension of the hidden state vector at each time point; Multi-scale feature extraction is performed by dividing the window into sliding windows. At each point in time Take all values ​​falling within the interval , Subset of hidden states ; right Perform self-attention pooling with positional encoding: in, This is a vector obtained through the hidden state; This is used to perform a dot product with the query Q vector to evaluate the matching degree; This is a value vector used to obtain the final output; , , These are the weight matrices for vectors Q, K, and V, respectively. This represents the weight vector of the attention mechanism. This is the final output vector.

4. The gas concentration prediction method according to claim 1, characterized in that, Step S3 specifically includes: designing a hybrid expert architecture, employing a two-way expert network, including a mining expert network and a maintenance expert network. The mining expert network is used to handle high-frequency, drastic fluctuation scenarios, while the maintenance expert network is used to handle low-frequency, slowly changing accumulation scenarios; constructing a gating network and outputting routing weights; and finally, the expert layer outputs a weighted sum feature.

5. The gas concentration prediction method according to claim 4, characterized in that, In step S3, each expert network is a multi-layer Transformer used to capture long-short-term dependencies and non-linear features.

6. The gas concentration prediction method according to claim 1, characterized in that, In step S5, the loss function of the prediction network is: in, To predict network loss, To predict the future of network output Predicted concentration values ​​at time [time]. The actual concentration value at time t_(n+1) For regularization parameters, For the first k Data collected by road sensors.

7. The gas concentration prediction method according to claim 1 or 6, characterized in that, In step S5, the prediction network uses LSTM, GRU, or Transformer networks.

Images