Double-flow gas concentration prediction method and device based on time sequence decomposition

Abstract

The invention discloses a double-flow gas concentration prediction method and device based on time sequence decomposition. The method comprises the steps that a multi-channel gas concentration data historical sequence of a first time period is obtained; decoupling the multi-channel gas concentration data historical sequence into a trend sub-sequence and a period sub-sequence based on the channels; obtaining a trend prediction sequence and a period prediction sequence of the gas concentration; fusing the trend prediction sequence and the period prediction sequence to obtain a predicted gas concentration sequence of the to-be-predicted time period; according to the method, the multi-channel gas concentration data historical sequence is decoupled into the trend sub-sequence and the periodic sub-sequence, so that the multi-scale feature capturing capability is improved, the modeling and fusion capability of time-varying interaction features among multiple channels can be enhanced, the adaptability in a complex environment is improved, and the prediction stability and robustness are improved.

Description

Technical Field

[0001] This invention belongs to the field of coal mine safety monitoring technology, and in particular relates to a method and device for predicting dual-flow gas concentration based on time-series decomposition. Background Technology

[0002] Gas concentration is a core indicator of coal mine safety, and its prediction accuracy directly affects the effectiveness of disaster prevention and control. Traditional monitoring methods mostly rely on single-point threshold exceeding alarms, lacking a forward-looking characterization of the risk evolution process. Although statistical models (such as ARIMA and SVM) can capture time-series patterns to some extent, they are difficult to effectively characterize complex nonlinear dynamic features.

[0003] Deep learning methods (such as LSTM, CNN, and Transformer) have been gradually applied to the field of gas concentration prediction, significantly improving the adaptive modeling capabilities of the models. However, when dealing with the non-stationarity of sequences, multi-scale coupling, and complex multivariate interactions in the underground environment, these methods still suffer from limitations in prediction accuracy, insufficient long-range dependency modeling capabilities, and high computational complexity. Therefore, in the face of strong noise disturbances and highly fluctuating sequences in the underground coal mine environment, the prediction accuracy drops significantly, making it difficult to meet the high-reliability safety early warning requirements. Summary of the Invention

[0004] The purpose of this invention is to provide a method and apparatus for predicting dual-flow gas concentration based on time-series decomposition, so as to improve prediction accuracy when facing strong noise disturbances and high fluctuation sequences.

[0005] This invention adopts the following technical solution: a method for predicting dual-flow gas concentration based on time-series decomposition, comprising the following steps: Obtain the historical sequence of multi-channel gas concentration data for the first time period; wherein, the first time period is the consecutive time period immediately preceding the time period to be predicted; the historical sequence of multi-channel gas concentration data includes at least one of the following: gas concentration in the fully mechanized mining face and the upper corner gas concentration, gas concentration in the return airway, temperature, gas concentration at the air inlet, wind speed, oxygen concentration, and coal mining machine cutting speed. Based on the channel, the historical sequence of multi-channel gas concentration data is decoupled into a trend subsequence and a periodic subsequence; the trend prediction sub-module is used to predict the trend subsequence to obtain the trend prediction sequence of gas concentration; the periodic prediction sub-module is used to predict the periodic subsequence to obtain the periodic prediction sequence of gas concentration; the trend prediction sequence and the periodic prediction sequence are fused to obtain the predicted gas concentration sequence for the time period to be predicted.

[0006] The beneficial effects of this invention are: by decoupling the historical sequence of multi-channel gas concentration data into trend sub-sequences and periodic sub-sequences, this invention improves the ability to capture multi-scale features, enhances the modeling and fusion capabilities of time-varying interactive features between multiple channels, improves adaptability in complex environments, and enhances prediction stability and robustness. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of a dual-flow gas concentration prediction method based on time-series decomposition according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the architecture of the dual-stream prediction module in an embodiment of the present invention; Figure 3 This is a schematic diagram comparing the effects of single-time-point prediction in an embodiment of the present invention; Figure 4 This is a schematic diagram comparing the effects of the model prediction with a sliding step size of 4 in an embodiment of the present invention; Figure 5 This is a schematic diagram comparing the effects of the model prediction with a sliding step size of 8 in an embodiment of the present invention; Figure 6 This is a schematic diagram comparing the effects of the model prediction with a sliding step size of 16 in an embodiment of the present invention. Detailed Implementation

[0008] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0009] Time series data on underground coal mine gas concentrations are highly suitable for analysis using time-series decomposition methods. The core reason is that the data characteristics perfectly match the applicable prerequisites of time-series decomposition, and the decomposed data can directly serve the actual production needs of gas control. This can be explained in detail from three dimensions: the inherent patterns in the data, the technical adaptability of the decomposition, and its practical application value.

[0010] I. The time series of gas concentrations possesses the core characteristics of time series decomposition.

[0011] The premise of time series decomposition is that the data can be decomposed into a superposition or product model of trend term + seasonal term + residual term, and gas concentration data naturally possesses these three components with clear boundaries: 1. There is a clear long-term trend.

[0012] Long-term variations in gas concentration are determined by geological conditions and mining activities, exhibiting strong trends that can be explained. As the working face advances, the gas content and permeability of the coal seam will show continuous changes (for example, when advancing to the gas enrichment zone, the concentration trend increases).

[0013] As the goaf expands, the amount of gas accumulated in the goaf increases, and the gas concentration in the return airflow will slowly rise. After long-term operation of the gas extraction system, changes in extraction efficiency will also lead to a trend of decreasing / increasing concentration. This trend is slow and continuous, and will not disappear due to short-term disturbances; it is the core basis of time-series decomposition.

[0014] 2. There are significant periodic seasonal terms.

[0015] The periodic fluctuations in gas concentration are entirely driven by the fixed production rhythm of the coal mine, with a clear and stable cycle, which is a key condition for time series decomposition: Daily cycle (24h): Coal mines generally adopt a "three-shift system" for production. During the production shift (early / mid / night shift), the operation of equipment and the work of personnel disturb the coal body, resulting in a large amount of gas emission and a high concentration. During the maintenance shift, the equipment is shut down, the coal body is stable, and the amount of gas emission is small and the concentration is low, forming a fixed 24-hour cycle.

[0016] Weekly cycle (168h): During normal production on weekdays, the gas concentration remains high; during weekend shutdowns for maintenance, the concentration drops significantly, forming a 7-day cycle fluctuation.

[0017] Monthly / Quarterly Cycles: Affected by mining plans and changes in geological structures, some mines may experience monthly-level cyclical fluctuations (such as concentration changes before and after crossing faults).

[0018] These cyclical fluctuations are "predictable and repeatable," which aligns with the definition of "seasonal term" in time series decomposition.

[0019] II. The technical characteristics of time-series decomposition are adapted to the analytical needs of gas concentration.

[0020] The core objective of coal mine gas concentration analysis is to "distinguish between controllable patterns and uncontrollable disturbances, and achieve accurate early warning and control," and the technical characteristics of time series decomposition can precisely meet this requirement: 1. Separate the "deterministic components" from the "random components" to avoid confusing patterns with noise.

[0021] In the raw data of gas concentration, trends and cycles are superimposed. For example, a rise in concentration at a certain moment could be due to a "trend-driven rise in mining progress," a "periodic rise in production shifts," or a "random rise due to ventilation fluctuations."

[0022] Directly analyzing raw data can easily lead to misinterpreting periodic fluctuations as trend changes (e.g., mistaking an increase in production shift concentration for impending gas enrichment), resulting in over-treatment. Conversely, using raw data directly for modeling and early warning can easily lead to misinterpreting random noise as abnormal exceedances (e.g., instantaneous high values ​​caused by sensor drift), triggering false alarms. Time series decomposition can completely separate these two aspects, allowing analysts to clearly identify "which changes are inevitable and which are accidental."

[0023] 2. It is highly robust and adaptable to "dirty data" in industrial scenarios.

[0024] The underground environment in coal mines is harsh, and gas concentration monitoring data inevitably contains problems such as outliers, missing values, and sensor drift (e.g., data loss due to instantaneous gas exceeding limits or equipment power failure).

[0025] Traditional statistical methods (such as linear regression) are sensitive to outliers, and a single outlier may cause the entire model to fail. The STL decomposition method in time series decomposition supports a "robust mode" that can automatically reduce the impact of outliers on trend and period extraction, ensuring the reliability of the decomposition results.

[0026] 3. No strict stationarity assumption is required, making it suitable for non-stationary gas data.

[0027] Many time series analysis methods (such as ARIMA) require the data to be stationary, while gas concentration data are typical non-stationary series (the trend term changes continuously).

[0028] Forcibly substituting non-stationary data into a stationary model will lead to large modeling errors and low prediction accuracy. Time series decomposition does not require the data to be stationary. Instead, it takes "non-stationary and periodic" as the premise of applicability and directly decomposes the structure of non-stationary sequences, which is more in line with the characteristics of gas data.

[0029] Third, the results of time-series decomposition directly serve the actual production of gas control.

[0030] The core requirements for coal mine gas management are "trend prediction, cycle control, and anomaly early warning," and the components of time-series decomposition precisely correspond to these three requirements, possessing extremely high practical value.

[0031] 1. Trends: Guiding long-term gas management planning.

[0032] If the trend term continues to rise, it indicates that the gas emission from the working face is increasing, and gas extraction needs to be strengthened in advance (e.g., increasing the number of extraction boreholes and extending the extraction time) and ventilation intensity needs to be increased (e.g., replacing with high-power fans). If the trend term continues to decline, it indicates that the current control measures are effective, and resource allocation can be appropriately optimized (e.g., reducing extraction power and saving energy). Changes in the trend term are the core basis for formulating long-term gas control plans.

[0033] 2. Seasonal items: guide precise regulation in different time periods.

[0034] Based on intraday seasonal cycles, "time-sharing ventilation scheduling" can be implemented: production shifts increase ventilation to ensure concentrations do not exceed limits; maintenance shifts reduce ventilation to save energy, balancing safety and economy. Based on weekly seasonal cycles, "gas management during weekend shutdowns" can be optimized: predicting weekend concentration decline trends in advance allows for adjustments to extraction parameters, avoiding excessive extraction and resource waste. Understanding seasonal patterns is key to achieving "refined gas management."

[0035] In summary, the three major characteristics of underground gas concentration in coal mines—"trend certainty, periodic stability"—perfectly match the core objectives of time-series decomposition: "component breakdown and pattern extraction."

[0036] Compared to directly analyzing raw data, time series decomposition can more accurately reveal the intrinsic changing patterns of gas concentration, while providing full-process support for gas control, including "long-term planning, medium-term regulation, and short-term early warning." This is the fundamental reason why time series decomposition is suitable for gas concentration analysis.

[0037] Existing gas concentration prediction technologies still have shortcomings in terms of adaptability and accuracy in complex downhole environments. Linear methods, represented by the DLinear model, simplify the time series structure and reduce computational complexity through seasonal-trend decomposition. However, because their feature extraction and mapping processes are based on fixed linear mapping, their fitting accuracy is insufficient when dealing with nonlinear fluctuation signals, making it difficult to accurately reflect the abrupt changes and dynamic changes in gas concentration.

[0038] Furthermore, this type of method lacks an effective modeling mechanism for the time-varying correlation between multiple sensor measurement points, making it difficult to achieve dynamic fusion and interactive learning of multivariate information. In downhole high-noise conditions, signals are easily interfered with, resulting in poor stability and noise resistance of the predicted output. Other deep learning models (such as CNN, LSTM, and Transformer) possess strong feature learning capabilities, but each still has its own technical limitations: CNNs have limited receptive fields, making it difficult to capture long-range dependency information; LSTM structures are constrained by gradient vanishing and serial computation mechanisms, making training time-consuming and difficult to achieve rapid iteration; while Transformer models have global feature extraction capabilities, they have high computational complexity and are sensitive to local perturbations.

[0039] This invention discloses a method for predicting dual-flow gas concentration based on time-series decomposition, such as... Figure 1As shown, the process includes the following steps: obtaining a historical sequence of multi-channel gas concentration data for a first time period; wherein, the first time period is the immediately preceding consecutive time period of the time period to be predicted; the historical sequence of multi-channel gas concentration data includes at least one of the following: gas concentration at the fully mechanized mining face and upper corner, gas concentration in the return airway, temperature, gas concentration at the air intake, wind speed, oxygen concentration, and coal mining machine cutting speed; decoupling the historical sequence of multi-channel gas concentration data into a trend sub-sequence and a periodic sub-sequence based on the channels; predicting the trend sub-sequence based on the trend prediction sub-module to obtain a trend prediction sequence of gas concentration; predicting the periodic sub-sequence based on the periodic prediction sub-module to obtain a periodic prediction sequence of gas concentration; and fusing the trend prediction sequence and the periodic prediction sequence to obtain a predicted gas concentration sequence for the time period to be predicted.

[0040] This invention constructs a dual-flow prediction network (De-STNet) based on seasonal-trend decomposition to achieve the following technical effects: 1) Improve the accuracy of multi-step gas concentration prediction, especially maintaining high reliability under non-stationary and highly fluctuating conditions; 2) Enhance the model's ability to model and fuse time-varying interaction features among multiple variables, improving adaptability in complex environments; 3) Achieve differentiated modeling and information fusion of long-term trends and short-term seasonal fluctuations, improving the ability to capture multi-scale features; 4) Reduce the model's sensitivity to noise, improving prediction stability and robustness, thereby achieving high-precision and stable prediction of underground gas concentration.

[0041] The predicted gas concentration sequence of this invention is the gas concentration (GM) of the fully mechanized mining face within a certain time period. The result can be a prediction at a single time point (sequence length is 1) or a prediction at multiple time points (sequence length is greater than 1).

[0042] Specifically, GM showed a significant positive correlation with the gas concentration in the upper corner (GU) and the gas concentration in the return airway (GR), revealing a strong spatial coupling among the three. This is mainly due to the influence of airflow transport: the gas released during coal mining migrates with the airflow to the upper corner and the return airway, where an accumulation zone easily forms, while a small amount of residual gas continues to emerge from the return airway.

[0043] GM is moderately positively correlated with temperature (T) and gas concentration at the intake (GI). The composition of the gas at the intake directly affects the gas concentration at the fully mechanized mining face, while temperature changes regulate the desorption and diffusion rates of gas, thereby affecting the gas release process.

[0044] In contrast, the correlation between GM and wind speed (WS) is weak, mainly because wind speed in this area is strictly controlled and its variation is limited, making it difficult to significantly affect gas concentration.

[0045] The cutting speed of a coal mining machine directly affects the crushing strength of the coal and the gas desorption rate. As the cutting speed increases, the amount of gas released per unit time increases significantly, leading to a rise in gas concentration at the working face. Simultaneously, under relatively stable ventilation conditions, the increased gas concentration reduces the volume fraction of oxygen, causing the oxygen concentration to decrease.

[0046] It should be noted that in this invention, the gas concentration at the fully mechanized mining face is used as input data, and at least one of the following is also used as input data: gas concentration at the upper corner, gas concentration in the return airway, temperature, gas concentration at the air intake, wind speed, oxygen concentration, and coal mining machine cutting speed; that is, at least two types of input data are used. The output data is the gas concentration at the fully mechanized mining face for the time period to be predicted.

[0047] In this invention, data preprocessing is required before decoupling the multi-channel gas concentration data historical sequence. This mainly includes: detecting and removing abnormal data in the multi-channel gas concentration data historical sequence; imputing missing data in the multi-channel gas concentration data historical sequence; eliminating high-frequency disturbance components in the multi-channel gas concentration data historical sequence; and removing gas concentration data historical sequences from weakly correlated channels in the multi-channel gas concentration data historical sequence.

[0048] Detection and removal of outlier data.

[0049] During gas concentration data acquisition, external factors such as high temperature and humidity underground, dust deposition, and electromagnetic interference can easily cause anomalies in the monitoring signals, including sudden changes, zero-point drift, and signal jitter. To ensure the validity and reliability of the data, it is necessary to detect and remove abnormal data. In this field, box plot methods, Laida's criterion method, adjacent difference threshold method, sliding window mean method, or Lagrange interpolation method can be used for this operation.

[0050] This invention employs box plots to detect and remove outlier data, setting the threshold range to three times the interquartile range (3IQR) to remove significant distortion and noise while retaining true characteristic information related to potential risk fluctuations. For detected outliers, nearest neighbor mean interpolation is further used for repair. Where t represents time. This represents the data at time ti. This represents the data at time t+i, where k is the threshold for the number of data points during interpolation. This represents the data after interpolation.

[0051] Imputation of missing data.

[0052] Due to channel transmission instability and data acquisition system response delay, some missing values ​​may appear in the original gas concentration observation data. To ensure the integrity and continuity of the data sequence, it is necessary to interpolate the missing data, and the methods described above can be used. When the number of missing data is small, the Lagrange interpolation method is used to fill in the missing data in this embodiment of the invention.

[0053] Regarding the elimination of high-frequency disturbance components (i.e., noise suppression).

[0054] Due to the influence of high-frequency electromagnetic interference and environmental noise in the well, the original gas concentration signal often contains high-frequency disturbance components, which leads to a decrease in the stability and accuracy of the model during the trend feature extraction stage. To improve signal quality, noise reduction methods such as Kalman filtering, VMD, and EMD can be used.

[0055] In this embodiment of the invention, wavelet threshold denoising method is used to reduce noise in the original gas concentration data: High-frequency and low-frequency components at different scales are obtained by wavelet decomposition (selecting the sym5 wavelet basis function and setting the decomposition level to 3). For high-frequency components, a fixed threshold strategy is used for compression: in, Let n be the noise standard deviation and n be the signal length. The wavelet threshold is related to the distribution of the input data.

[0056] To avoid the oversmoothing problem caused by traditional soft thresholding functions and the Gibbs phenomenon caused by hard thresholding functions, the Garrote thresholding function is used to adjust the original wavelet coefficients ( Nonlinear compression is performed to enhance noise reduction in the threshold neighborhood while preserving the main dynamic structure of the signal. in, These are the original wavelet coefficients, where j and k represent the scale and location, respectively. For threshold parameters, For sign functions. When When greater than 0, +1, when When equal to 0, When it is 0, When less than 0, It is -1.

[0057] Finally, inverse wavelet transform is performed on the coefficients after thresholding to reconstruct the denoised signal.

[0058] Regarding channel correlation analysis.

[0059] To enhance the information effectiveness of input features, the Pearson correlation coefficient method is used to quantitatively analyze the correlation between gas concentration and multiple influencing factors in the fully mechanized mining face. By calculating the linear correlation coefficient between each influencing factor and gas concentration, feature variables with strong correlation to the target variable are selected, while feature inputs with low correlation or redundancy are eliminated. This reduces model dimensionality, weakens noise interference, and improves the effectiveness of feature representation and model prediction accuracy. Specifically, the Pearson correlation coefficient method includes: Where r is the Pearson correlation coefficient, and X and Y represent the sample values ​​of the two variables, respectively. and is the sample mean, and n is the sample size.

[0060] Next, the gas concentration is predicted using a dual-flow prediction module based on seasonal-trend decomposition.

[0061] Trend prediction submodule input: original gas concentration sequence X∈ C represents the number of channels, and L represents the length of the historical sequence; Output: Trend subsequence X T ∈ With periodic subsequence X S ∈ The specific method is as follows: AvgPool is a moving average pooling method with padding, used to extract low-frequency trends. Indicates a concatenation operation; periodic subsequence It represents the difference between the original sequence and the trend term, reflecting high-frequency fluctuations.

[0062] The core of the above decoupling (i.e., time series decomposition) lies in decoupling the non-stationary original gas concentration sequence into components at different time scales to reduce modeling complexity. Of course, in addition to seasonal-trend decomposition (STD), exponential moving average (EMA), empirical mode decomposition (EMD), or variational mode decomposition (VMD) can also be used to achieve feature decoupling at different time scales.

[0063] The trend prediction submodule performs reversible normalization and denormalization on the trend subsequence corresponding to each channel; there is a linear subnetwork between the normalization and denormalization, and the linear subnetwork has several cascaded fully connected layers.

[0064] In one embodiment, such as Figure 2 As shown, Reversible Instance Normalization (RevIN) is used: each sample is normalized independently to alleviate distribution drift.

[0065] Normalization: Input is (Sequence length is L), the output is : in, and The mean and standard deviation of the samples within each channel dimension. To be the minimum value, guarantee Not zero.

[0066] Inverse normalization: The prediction result of the trend prediction sequence at time t as input. trend The output is :

[0067] Channel Independent MLP: A separate linear subnetwork is constructed for each channel (variable) in a multivariate time series for prediction, avoiding potential information interference and noise coupling between variables.

[0068] The linear subnetwork consists of a first fully connected layer, a second fully connected layer, and a third fully connected layer. For the normalized trend subsequence... t ∈ The first fully connected layer is used to map the normalized trend subsequence to a high-dimensional feature space, resulting in... t ∈ , Let C represent a real number space with C channels and D sequence length. t Represents the high-dimensional features at time t; the second fully connected layer is used to... t Mapping to the prediction space, generating trend predictions for each channel, yielding... t ∈ , Let T represent a real space with C channels and T sequence length. t This represents the channel prediction characteristics at time t; Finally, the third fully connected layer is used to weight and integrate the outputs of each channel along the channel dimension, t Mapping to a single-channel prediction space yields trend ∈ , Let T represent the real space of a sequence of length T in a predetermined channel. trendThis represents the trend prediction sequence at time t.

[0069] In addition to channel-independent multilayer perceptrons (MLPs), long short-term memory networks or gated recurrent units can also be used. Their recursive structures can effectively model time dependencies and capture long-term evolution patterns contained in trend terms. Alternatively, the Transformer encoder structure can be used to directly model the dependencies of all time steps within the sequence using its global attention mechanism, which is good at capturing complex long-term correlation patterns.

[0070] In this invention, the method of predicting gas concentration based on the periodic prediction sub-module to obtain the periodic prediction sequence includes: reconstructing the normalized periodic sub-sequence into a patch sequence; performing depthwise separable convolution and dynamic convolution on the patch sequence in sequence to obtain an integrated feature map; mapping the integrated feature map and the patch sequence to the prediction space after residual connection; and fusing the feature maps of the prediction space of each channel to obtain the periodic prediction sequence.

[0071] Patching: Divides a sequence into subsequences to improve long-range modeling capabilities.

[0072]

[0073] Where L is the time series length, P is the patch length, S is the sliding step size, and N is the continuous subsequence after partitioning.

[0074] Periodic subsequences t ∈ Reconstructed into a patch sequence Where D is the feature dimension of each patch and N is the number of patches.

[0075] Depthwise Convolution: Performs temporal convolutions independently on each variable to generate channel-independent feature representations. This involves extracting local temporal features. Specifically, depthwise separable convolutions are used to independently model each variable in the time dimension, thereby extracting the dynamic evolution patterns of individual variables and avoiding information interference caused by channel coupling.

[0076] Dynamic convolution is achieved through a first dynamic convolutional layer and a second dynamic convolutional layer connected in series.

[0077] Dynamic convolution: Based on an attention mechanism, it dynamically adjusts the weights of convolution kernels to enhance the modeling ability of time-varying interactions among multiple variables. Specifically, the dynamic convolution module adaptively combines the weights of multiple 1×1 convolution kernels, allowing the kernel parameters to be dynamically adjusted according to sample features and time steps. This more effectively captures the nonlinear dependencies and dynamic features between channels, improving the model's ability to express complex gas time-series fluctuations.

[0078] The first dynamic convolutional layer is used to capture single-channel nonlinear temporal variations and outputs the first integrated feature map. , Let C represent a real space with C channels, D sequence length, and N patch sequences; that is, DynamicConv1 adopts a group = C structure for intra-variable modeling, i.e., it independently generates dynamic convolutional kernels for each channel to capture univariate nonlinear temporal variations and outputs... .

[0079] The second dynamic convolutional layer is used to achieve integrated representation of multi-channel information and outputs a second integrated feature map. In other words, DynamicConv2 uses a group = D (high-dimensional space) structure to model the interactions between variables. This involves dynamically convolving inter-channel features to learn the latent interaction patterns between different variables, achieving an integrated representation of multi-variable information, and outputting the result. .

[0080] Furthermore, a fourth fully connected layer maps the residually connected feature maps to the prediction space. Residual connections are used to improve the training stability of deep networks and alleviate the vanishing gradient problem. The residually connected features are then mapped to the prediction space. The system generates seasonal predictions for each channel; finally, it uses a linear fusion module to weight and integrate the outputs of each channel along the channel dimension to obtain the seasonal prediction results. .

[0081] In addition to combinations of depthwise separable convolutions and dynamic convolutions, attention networks incorporating self-attention or multi-head attention mechanisms can be used to dynamically calculate the importance of different time steps, accurately capturing key abrupt changes and periodic patterns in seasonal terms. Alternatively, temporal convolutional networks (TCNs) can be used to process high-frequency fluctuating data and extract local temporal dependent features.

[0082] Finally, the trend prediction sequence and the periodic prediction sequence are fused. The inversely normalized linear stream output and the inversely normalized nonlinear stream output are linearly fused to generate the final multi-step prediction result. The specific method is as follows: in, This represents a predicted gas concentration sequence. express trend Trend prediction sequence after inverse normalization This represents the periodic prediction sequence after inverse normalization, where α and β represent adjustment factors used to adaptively adjust the contribution ratio.

[0083] In summary, this invention employs a dual-flow decomposition mechanism to decompose the gas concentration time series into trend and seasonal terms, and constructs independent sub-networks for targeted modeling, thereby achieving structural decoupling and differentiated prediction of multi-scale features.

[0084] Linear Flow (Trend Term Prediction): By dynamically aligning the statistical distributions of different samples through reversible instance normalization, the problem of data distribution drift is significantly alleviated; and by using a channel-independent linear model, redundant interference among multiple variables can be reduced, effectively extracting long-term trend information.

[0085] Nonlinear Flow (Seasonal Term Prediction): Independent modeling of each variable in the time dimension is achieved through depthwise separable convolution, extracting the dynamic evolution law of single variables and avoiding feature interference caused by channel coupling; dynamic convolution adaptively adjusts the weight of convolution kernel to achieve dynamic fusion of time-varying interaction relationships among multiple variables, enhancing the model's ability to express complex nonlinear features.

[0086] Fusion prediction: By linearly fusing the prediction results of the trend term and the seasonal term, the prediction results are jointly output, taking into account both long-term trend stability and short-term fluctuation responsiveness, so as to achieve high-precision, multi-step prediction of gas concentration series.

[0087] This invention, through a general strategy of "temporal decomposition-differentiated modeling-linear fusion," addresses the non-stationarity, multi-scale temporal coupling, and complex multivariate correlations of gas concentration sequences in underground coal mine environments, achieving a significant improvement in prediction accuracy and stability. The method first uses seasonal-trend decomposition to structurally decompose the original gas concentration sequence, decoupling long-term trends from short-term fluctuations to mitigate the impact of non-stationary features on model training. Then, a differentiated modeling strategy is employed for different features: the trend term uses reversible instance normalization and channel-independent linear modeling to align sample distributions and characterize long-term evolution; the seasonal term uses depthwise separable convolution and dynamic convolution structures to extract local temporal features while adaptively modeling time-varying interactions among multiple variables and enhancing nonlinear expressive power. Finally, the two prediction results are linearly fused to meet the modeling needs of both long-term trends and short-term fluctuations.

[0088] To better illustrate the effectiveness of the method of this invention, a verification experiment was conducted between this method and existing prediction methods. In this experiment, five key features were selected as input variables for the prediction model: gas concentration at the fully mechanized mining face (GM), gas concentration in the return airway (GR), gas concentration at the upper corner (GU), temperature (T), and gas concentration at the air inlet (GI). This was to fully characterize the spatial coupling effect and environmental driving factors in the gas dynamics evolution process, thereby improving the prediction model's ability to characterize risk time series and its prediction accuracy.

[0089] When partitioning the dataset, the preprocessed gas concentration-related data were divided into training and test sets in an 8:2 ratio. A total of 7200 gas data points were collected, resulting in a training set of 5760 data points and a test set of 1440 data points. During model training, the batch size was set to 128, and the epoch was set to 100. The epoch with the lowest loss was selected for testing. The experimental environment is shown in Table 1.

[0090] Table 1 To verify the prediction performance of the gas concentration prediction method based on trend-seasonal decomposition, mean absolute error (MAE) and root mean square error (RMSE) were selected to evaluate the model performance. The smaller the index, the better the model prediction performance. The principles of the two indices are shown in equations (11) and (12): Where n represents the number of samples. Represents the actual value. This represents the predicted value.

[0091] To verify the predictive performance of the proposed model, comparative experiments were conducted using DLinear, ModernTCN, and PatchTST. All comparison models were trained and tested on the same training and test sets. To better demonstrate the performance of the prediction model on the test set, only partial test set results are shown. Considering both the real-time nature of the prediction results and long-term planning, the historical review window length was set to 120, and prediction step sizes were set to 1, 4, 8, and 12, respectively.

[0092] The evaluation metrics of each model under different prediction step sizes are shown in Table 2, and the prediction results of gas concentration for each model under different prediction step sizes are shown in Table 2. Figures 3-6 As shown.

[0093] Table 2 Table 2 shows that the De-STNet model has the best prediction results, with average MAE and RMSE of 0.0114 and 0.0147, respectively. Compared with the DLinear model, the average MAE and RMSE are reduced by 16.4% and 19.6%, respectively.

[0094] This invention allows setting the number of future prediction steps, categorized into single-step prediction (single-point prediction) and multi-step prediction (sequence prediction). During multi-step prediction visualization, due to the sliding window step size being set to 1, there is significant temporal overlap between adjacent prediction windows. To avoid multiple predicted values ​​corresponding to the same time point, only the first predicted value from each multi-step prediction result is selected during the visualization stage, thereby constructing a prediction curve that corresponds one-to-one with the actual sequence timeline. It should be noted that this processing is only for result display; model training and performance evaluation are still based on the complete multi-step prediction output. Figure 3 The image shows a comparison of the effects when performing single-time-point prediction. Figure 4 The image shows a comparison of the model's performance when the sliding step size is 4 during prediction. Figure 5 The image shows a comparison of the model's performance when the sliding step size is 8 during prediction. Figure 6 The image shows a comparison of the model's performance when the sliding step size is 16 during prediction. (The image is presented in a diagram.) Figures 3-6 It can be seen that the prediction results of the method of the present invention are closer to the actual results, thus improving the prediction accuracy.

[0095] The present invention also discloses a method for predicting dual-flow gas concentration based on time-series decomposition, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor implements the above-described method when executing the computer program.

[0096] The present invention also discloses an embodiment that provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0097] The present invention also provides a computer program product that, when run on a data storage device, enables the data storage device to implement the steps in the above-described method embodiments.

[0098] If the integrated unit module is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a storage device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0099] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0100] Those skilled in the art will recognize that the algorithmic steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0101] It should be noted that the data used in the implementation of this invention were all collected or gathered through legal and compliant channels, and the collection and gathering activities fully comply with the requirements of relevant laws, regulations and industry standards; the existing technical methods involved in this invention were also obtained and used through legal and compliant means.

Claims

1. A method for predicting dual-flow gas concentration based on time-series decomposition, characterized in that, Includes the following steps: Obtain a historical sequence of multi-channel gas concentration data for a first time period; wherein, the first time period is the consecutive time period immediately preceding the time period to be predicted; the historical sequence of multi-channel gas concentration data includes at least one of the following: gas concentration at the fully mechanized face and the upper corner, gas concentration in the return airway, temperature, gas concentration at the air inlet, wind speed, oxygen concentration, and coal mining machine cutting speed. Based on the channel, the historical sequence of multi-channel gas concentration data is decoupled into a trend subsequence and a periodic subsequence; the trend prediction sub-module predicts the trend subsequence to obtain a trend prediction sequence of gas concentration; the periodic prediction sub-module predicts the periodic subsequence to obtain a periodic prediction sequence of gas concentration; the trend prediction sequence and the periodic prediction sequence are fused to obtain a predicted gas concentration sequence for the time period to be predicted.

2. The method for predicting dual-flow gas concentration based on time-series decomposition as described in claim 1, characterized in that, The trend prediction submodule performs reversible normalization and denormalization on the trend subsequence corresponding to each channel; wherein, there is a linear subnetwork between the normalization and denormalization, and the linear subnetwork has several cascaded fully connected layers.

3. The method for predicting dual-flow gas concentration based on time-series decomposition as described in claim 1, characterized in that, The linear subnetwork includes a first fully connected layer, a second fully connected layer, and a third fully connected layer; The first fully connected layer is used to map the normalized trend subsequence to a high-dimensional feature space, resulting in... t ∈ , Let C represent a real number space with C channels and D sequence length. t Represents the high-dimensional features at time t; The second fully connected layer is used to... t Mapping to the prediction space, we get t ∈ , Let T represent a real space with C channels and T sequence length. t This represents the channel prediction characteristics at time t; The third fully connected layer is used to... t Mapping to a single-channel prediction space yields trend ∈ , Let T represent the real space of a sequence of length T in a predetermined channel. trend This represents the trend prediction sequence at time t.

4. A method for predicting dual-flow gas concentration based on time-series decomposition as described in claim 2 or 3, characterized in that, The periodic prediction sequence of gas concentration obtained by predicting the periodic subsequence based on the periodic prediction submodule includes: The normalized periodic subsequence is reconstructed into a patch sequence; The patch sequence is then subjected to depthwise separable convolution and dynamic convolution sequentially to obtain an integrated feature map; The integrated features and patch sequences are residually concatenated and mapped to the prediction space. The feature maps of the prediction spaces of each channel are then fused to obtain the periodic prediction sequence.

5. The method for predicting dual-flow gas concentration based on time-series decomposition as described in claim 4, characterized in that, The dynamic convolution is achieved through a first dynamic convolutional layer and a second dynamic convolutional layer connected in series; The first dynamic convolutional layer is used to capture single-channel nonlinear temporal variations and outputs a first integrated feature map. , Let N be the real number space containing C channels, D sequence lengths, and N patch sequences. The second dynamic convolutional layer is used to achieve integrated representation of multi-channel information and output a second integrated feature map. .

6. The method for predicting dual-flow gas concentration based on time-series decomposition as described in claim 5, characterized in that, The feature map after residual connection is mapped to the prediction space through the fourth fully connected layer; The feature maps of the prediction space of each channel are fused through the fifth fully connected layer to obtain the periodic prediction sequence.

7. The method for predicting dual-flow gas concentration based on time-series decomposition as described in claim 6, characterized in that, The method for fusing the trend prediction sequence and the period prediction sequence is as follows: ， in, This represents a predicted gas concentration sequence. express trend Trend prediction sequence after inverse normalization This represents the periodic prediction sequence after inverse normalization, where α and β represent adjustment factors used to adaptively adjust the contribution ratio.

8. The method for predicting dual-flow gas concentration based on time-series decomposition as described in claim 2, characterized in that, Before decoupling the historical sequence of the multi-channel gas concentration data, the following steps are also included: Abnormal data in the historical sequence of the multi-channel gas concentration data are detected and removed; Missing data in the historical sequence of the multi-channel gas concentration data are imputed; Eliminate high-frequency disturbance components in the historical sequence of the multi-channel gas concentration data; Remove the weakly correlated channels from the historical gas concentration data sequence of the multi-channel gas concentration data.

9. A method for predicting dual-flow gas concentration based on time-series decomposition, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1-8.

Images