Method for predicting gas concentration of working face for long time based on DTW-BiT-Times Net model

Abstract

The invention relates to a working face gas concentration long-time prediction method based on a DTW-BiT-Times Net model, and belongs to the technical field of coal mine safety monitoring. The method comprises the following steps: collecting monitoring data of gas, air speed and dust sensors of a working face and an air return site for more than three continuous months, and extracting sub-mean values to form a data column; data preprocessing is completed through verification data filtering, time-lag correlation analysis and anomaly correction; and constructing a DTW-BiT-Times Net model, fusing first k frequency corresponding periods, a mine shift period and a cyclic operation period of frequency domain analysis, iteratively analyzing multi-dimensional time sequence characteristics by utilizing a weighted summation formula, and performing model training and verification by adopting a Soft-DTW loss function to realize long-time prediction of the gas concentration of the working face. The problems of error accumulation, feature loss and insufficient robustness are solved through multi-sensor data association verification, dynamic anomaly correction and periodic feature embedding.

Description

Technical Field

[0001] The present invention belongs to the technical field of coal mine safety monitoring, and relates to a method for long-term prediction of gas concentration in the working face based on the DTW-BiT-TimesNet model. Background Technique

[0002] Mine gas is one of the main threats to coal mine safety production. Abnormal gas concentration is extremely likely to trigger major accidents such as coal and gas outbursts, fires or explosions. Real-time monitoring and accurate prediction of gas concentration in the working face are the core basis for disaster early warning. However, the coal mine working face changes dynamically with the progress of mining and excavation, while the air return location remains fixed. The gas concentration is affected by the coupling of multiple factors such as geological conditions, ventilation efficiency, and equipment operation, showing highly non-linear and time-varying characteristics, which poses a severe challenge to long-term prediction technology.

[0003] Existing gas concentration prediction methods mainly rely on statistical models (such as ARIMA), seasonal trend decomposition models (STL), and deep learning models (such as LSTM, GRU, Autoformer). Although these methods have achieved certain results in short-term prediction, their single-step iterative prediction mode is prone to error accumulation and difficult to meet the long-term prediction requirements. In addition, the existing technologies have the following limitations:

[0004] (1) Existing methods usually adopt conventional operations such as standardization and null value filling, but lack effective processing of gas sensor calibration anomalies (such as calibration drift, instantaneous noise), resulting in low data quality;

[0005] (2) Existing models do not fully exploit the spatio-temporal coupling relationship between gas sensors in the working face and sensors at the air return location (such as wind speed, dust sensors), and key feature extraction is insufficient, which is prone to model overfitting;

[0006] (3) Mine operations have significant shift cycles and cyclic operation cycles, but existing models do not specifically embed such time series features, resulting in a deviation between the prediction results and the actual working conditions;

[0007] (4) Existing methods mostly rely on statistical thresholds (such as the 3σ statistical anomaly recognition method) to identify anomalies, do not consider the data feature requirements of gas disaster analysis, and do not combine the sensor time delay differences for joint identification, resulting in a high misjudgment rate.

[0008] In view of the above problems, there is an urgent need for a long-term prediction method that integrates multi-source sensor data, combines the characteristics of mine operation cycles, and has a high-robustness anomaly processing ability to achieve early and accurate warning of gas concentration in the working face and provide reliable technical support for the coal mine intelligent safety system. Summary of the Invention

[0009] In view of this, the purpose of the present invention is to provide a long-term prediction method for the gas concentration of the working face based on the DTW-BiT-TimesNet model.

[0010] To achieve the above object, the present invention provides the following technical solutions:

[0011] A long-term prediction method for the gas concentration of the working face based on the DTW-BiT-TimesNet model, comprising the following steps:

[0012] S1: Collect the monitoring data of the gas sensor, return air sensor, wind speed sensor, and dust sensor at the working face location in the coal mine safety monitoring system for no less than 3 months under continuous and stable mining conditions, extract the average value and record it as data column A i , i ∈ [1, 4];

[0013] S2: Perform calibration data filtering and filling on data columns A1 and A2, and perform anomaly identification and correction based on time lag and correlation; perform data standardization on data columns A1, A2, A3, and A4, generate a basic data set, and divide it into a training set, a test set, and a validation set according to a ratio of 7:2:1;

[0014] S3: Take A1, A2, A3, and A4 as input quantities, and A1 as the predicted quantity. Based on the mine shift cycle T bc and the cyclic operation cycle T xh Construct a DTW-BiT-TimesNe model, use the gradient descent method and the Soft-DTW loss function for model training and verification, and output the long-term prediction result of the gas concentration A1 of the working face.

[0015] Further, the calibration data filtering includes:

[0016] Identify abnormal data according to the regular calibration rules of the mine, set the calibration time interval [t a , t b and the calibration value K, traverse this interval to calculate the absolute value of the difference v j , j ∈ N + , when two consecutive difference values are greater than the specified threshold and the monitoring value exceeds 0.4K, mark it as an abnormal point, and extend the abnormal point interval to the left and right safety thresholds m. Finally, fill the data in the abnormal interval through the adjacent interpolation algorithm.

[0017] Further, the anomaly identification and correction based on time lag and correlation include:

[0018] First-order difference operations are performed on A1 and A2 to obtain B1 and B2. Moving averages with a window of w are performed on A1 and A2, and then shifted forward by w / 2 positions to fill in the blanks, resulting in C1 and C2, where w is an even number. Based on the 3σ statistical anomaly recognition method, after identifying the differential values at abnormal time points in B1 and B2, the parts with differential values greater than C1 and C2 are selected and determined as abnormal pending time points T c1 and T c2 . Combining the time delay Δt between the working face and the return air sensor, it is judged whether the abnormal points are paired. If the abnormal points are paired, they are determined as real anomalies and retained; otherwise, they are determined as local anomalies, deleted, and interpolated to fill in the blanks.

[0019] Furthermore, the construction of the DTW-BiT-TimesNe model includes:

[0020] Perform frequency domain analysis on the input data and select the k periods T corresponding to the k highest frequency values i , i ∈ [1, k], and the mine shift cycle T bc and the cyclic operation cycle T xh The data is sliced into k + 2 two-dimensional data with the length of the period value. After extracting features by convolution of the k + 2 two-dimensional arrays based on the Inception Block in TimesNet and then converting them back to k + 2 one-dimensional arrays, the time series features are analyzed layer by layer through the following weighted summation formula:

[0021]

[0022] In the formula, l is the number of layers of the time series analysis block, is the output of the (l - 1)-th layer time series analysis block and the input of the l-th layer time series analysis block; is the data after analysis and conversion to one dimension of the (l - 1)-th layer time series analysis block; Concat(·) is the merging of one-dimensional arrays; Softmax(·) is the normalized exponential function; is the amplitude corresponding to the Top K frequencies after frequency domain decomposition of the (l - 1)-th layer; w bc , w xh are the undetermined weight parameters corresponding to the decomposition of the mine shift cycle and the cyclic operation cycle respectively.

[0023] Furthermore, the Soft-DTW loss function is defined as:

[0024]

[0025] In the formula, y * is the actual value, is the predicted value, is the initialized cost matrix, solved with reference to the DTW algorithm; is the inner product of the alignment matrix A and the cost matrix; min γ{·} is a differentiable minimization function containing the hyperparameter γ, defined as:

[0026]

[0027] where a i , i ∈ [1, n] are the terms for which the minimum value needs to be found, and n is the number of terms.

[0028] Furthermore, in S2, data standardization includes: performing standardization processing on A1, A2, A3, and A4 to generate a basic data set with a unified scale after eliminating the dimension difference.

[0029] Furthermore, the length of the verification time interval is one day, i.e., t b -t a = 1440 min, and the left and right safety thresholds m are taken as 5 min.

[0030] Furthermore, the value range of the hyperparameter γ of the Soft-DTW loss function is 0 to 1.

[0031] Furthermore, the selection criterion for the Top K frequencies is the first 5 frequencies arranged in descending order of amplitude.

[0032] Furthermore, during the model training, the dynamic adjustment range of the learning rate of the gradient descent method is 10 -5 ~10 -3 .

[0033] The beneficial effects of the present invention are as follows:

[0034] (1) By verifying and processing the sensor data, and using the sensor data related to the working face, the present invention can effectively improve the accuracy of long-term prediction of gas concentration and provide more reliable data support for early warning of disasters.

[0035] (2) Using the DTW-BiT-TimesNet model for long-term prediction can effectively avoid the cumulative error caused by single-step iteration and improve the stability of the prediction results.

[0036] (3) The present invention takes into account the influence of the mine shift cycle and the cyclic operation cycle on the gas concentration, and uses the sensor data of the working face and the return air location, which can more comprehensively reflect the change law of the gas concentration of the working face.

[0037] (4) Through data standardization and model training optimization, the present invention can improve the prediction efficiency and meet the requirements of real-time monitoring and early warning.

[0038] Improving the safety production level of coal mines: The present invention can provide technical support for early warning of coal and gas outbursts, fire early warning, and gas explosion early warning, and effectively improve the safety production level of coal mines.

[0039] Other advantages, objectives and features of the present invention will, to some extent, be described in the following specification, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the following specification. Brief Description of the Drawings

[0040] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in detail preferably with reference to the accompanying drawings, where:

[0041] Figure 1 It is a long-term prediction flowchart of gas concentration in a mine working face;

[0042] Figure 2 It is a schematic diagram of the principle of checking data noise filtering;

[0043] Figure 3 It is the abnormal recognition and processing based on time lag and correlation;

[0044] Figure 4 It is the DTW-BiT-TimesNet iterative time series analysis block structure;

[0045] Figure 5 It is the average monitoring data of four sensors at the driving working face of a certain mine in Guizhou;

[0046] Figure 6 It is the result diagram of checking data filtering and filling for the gas sensor data A1 at the working face;

[0047] Figure 7 It is the result diagram of checking data filtering and filling for the return air gas sensor data A2;

[0048] Figure 8 It is the result diagram of abnormal recognition and correction for the gas sensor data A1 at the working face;

[0049] Figure 9 It is the result diagram of abnormal recognition and correction for the return air gas sensor data A2;

[0050] Figure 10 It is the prediction comparison diagram of four groups of gas sensor data A1 at the working face. Detailed Description of the Preferred Embodiments

[0051] The following describes the embodiments of the present invention through specific examples. Those skilled in the art can easily understand the 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. The details in this specification can also 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 drawings provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0052] Among them, the drawings are only for illustrative purposes, showing only schematic diagrams, not physical diagrams, and should not be construed as a limitation to the present invention; in order to better illustrate the embodiments of the present invention, some components in the drawings will be omitted, enlarged or reduced, which does not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0053] In the drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "rear", etc. indicating the orientation or positional relationship, they are based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the drawings are only for illustrative purposes and should not be construed as a limitation to the present invention. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0054] As Figure 1 shown, the present invention provides a method for long-term prediction of gas concentration in a mine working face, including the following steps:

[0055] S1: Collect the monitoring data of the gas sensor, return air sensor, wind speed sensor and dust sensor at the working face location and the return air location under continuous and stable mining and excavation conditions for no less than 3 months from the coal mine safety monitoring system, and extract the average values, which are respectively recorded as data columns A i , i ∈ [1, 4];

[0056] S2: First, perform calibration data filtering and filling on A1 and A2, and then perform abnormal identification and correction based on time lag and correlation; then standardize the data of A1, A2, A3, and A4 to generate a basic data set; finally, divide the training set, test set and validation set according to 7:2:1.

[0057] S3: Set A1, A2, A3, and A4 as input quantities, and A1 as the predicted quantity, based on the mine shift cycle T bcand the cyclic operation period T xh Embed and construct the DTW-BiT-TimesNe model, and use the gradient descent method and the Soft-DTW loss function to train and verify the model to achieve long-term prediction of the gas concentration A1.

[0058] As Figure 2 shown, for the calibration data filtering described in S2, first, identify the abnormal data formed by calibration through the rules of regular calibration and fixed-value calibration in the mine. Set a certain calibration time interval as [t a , t b , and the calibration calibration value is K. Then traverse the time interval to perform data difference and take the absolute value to get v j , j ∈ N + . When two consecutive difference values are greater than the specified value and the monitored value is greater than 0.5K, count it as an abnormal point, and expand the abnormal point interval by the specified safety threshold m to the left and right to obtain the data calibration interval; then, fill it through the nearest neighbor interpolation algorithm.

[0059] As Figure 3 shown, for the abnormal identification and correction based on time lag and correlation described in S2, first, perform first-order difference operations on A1 and A2 to obtain B1 and B2, perform moving average with a window of w on A1 and A2, and then move forward by w / 2 bits to fill in to obtain C1 and C2, where w is an even number; based on the 3σ statistical anomaly identification method, identify the abnormal time point difference values in B1 and B2, and then select the part where the difference value is greater than C1 and C2, and determine it as the abnormal pending time point T c1 、T c2 . Combine the time delay Δt of the working face and the return air sensor to judge whether the abnormal points are paired; if the abnormal points are paired, judge them as real abnormalities and retain them; otherwise, judge them as local abnormalities, delete them, and fill them by interpolation.

[0060] As Figure 4 shown, for the DTW-BiT-TimesNet model described in S3, the data decomposition therein is to select the Top K frequencies corresponding to the period T i , i ∈ [1, k], as well as the mine shift period T bc and the cyclic operation period T xh , perform data segmentation and merge them into k + 2 groups of two-dimensional data. The weighted summation formula is as follows:

[0061]

[0062] In the formula, is the output of the l-1 layer time series analysis block; is the data after the l-1 layer time series analysis block analyzes and converts to one dimension; Concat(·) is the merging of one-dimensional arrays; Softmax(·) is the normalization exponential function; is the amplitude corresponding to the Top K frequencies after the l-1 layer frequency domain decomposition; here, K = 5; w bc , w xh are the undetermined weight parameters corresponding to the decomposition of the mine shift cycle and the cyclic operation cycle respectively.

[0063] The Soft-DTW loss function described in S3 is the actual value y of length n * and the predicted value to calculate the differentiable implementation of the dynamic time warping distance function DTW, and the formula is as follows

[0064]

[0065] In the formula, is the initialized cost matrix, solved with reference to the DTW algorithm; is the inner product of the alignment matrix A and the cost matrix; min γ {·} is a differentiable minimization function containing the hyperparameter γ, defined as

[0066]

[0067] As Figure 5 shown, the average data of the driving monitoring for 91 days and 131,040 minutes from September 4, 2022 to December 4, 2022 in the transportation roadway of the 1501 working face of the C5 coal seam in a certain mine was collected, including the data of the working face gas sensor A1, the return air gas sensor A2, the return air wind speed sensor A3, and the return air dust sensor A4. The data statistical information is shown in Table 1.

[0068] Table 1

[0069] Data Name Unit Maximum Value Minimum Value Average Value Coefficient of Variation <![CDATA[Gas sensor data A1 at the working face]]> ％ 1.58 0.00 0.08 0.60 <![CDATA[Return air gas sensor data A2]]> ％ 1.52 0.04 0.20 0.47 <![CDATA[Return air velocity sensor data A3]]> m / s 0.90 0.30 0.56 0.21 <![CDATA[Return air dust sensor data A4]]> <![CDATA[mg / m 3 > 675.23 0.00 5.10 3.67

[0070] As Figure 6 , Figure 7 shown, the mine carried out data calibration on September 15, 2022, October 1, 2022, October 16, 2022, November 1, 2022, and November 15, 2022. The calibration time intervals are [September 15, 2022 00:00:00, September 15, 2022 23:59:59], [October 1, 2022 00:00:00, October 1, 2022 23:59:59], [October 16, 2022 00:00:00, October 16, 2022 23:59:59], [November 1, 2022

[0071] 00:00:00, November 1, 2022 23:59:59], [November 15, 2022 00:00:00, November 15, 2022 23:59:59].

[0072] Taking the calibration of the working face gas sensor A1 on September 15, 2022 as an example to illustrate the filtering of calibration data, Table 2 shows the data for 20 minutes before and after calibration.

[0073] Table 2

[0074] Serial Number Time Point Value Serial Number Time Point Value Serial Number Time Point Value 1 12:50 0.08 8 12:56 0.06 15 13:04 0.04 2 12:51 0.09 9 12:57 0.74 16 13:05 0.04 3 12:52 0.09 10 12:58 1.40 17 13:06 0.04 4 12:53 0.09 11 12:59 1.54 18 13:07 0.03 5 12:54 0.09 12 13:00 1.12 19 13:08 0.03 6 12:55 0.09 13 13:01 0.11 20 13:09 0.04 7 12:56 0.09 14 13:03 0.04 21 13:10 0.04

[0075] Perform a difference operation on the data and take the absolute value to obtain Table 3.

[0076] Table 3

[0077] Serial Number Time Point Value Serial Number Time Point Value Serial Number Time Point Value 1 12:50 - 8 12:56 0.03 15 13:04 0.00 2 12:51 0.01 9 12:57 0.67 16 13:05 0.00 3 12:52 0.00 10 12:58 0.67 17 13:06 0.00 4 12:53 0.00 11 12:59 0.14 18 13:07 0.01 5 12:54 0.00 12 13:00 0.41 19 13:08 0.00 6 12:55 0.00 13 13:01 1.01 20 13:09 0.01 7 12:56 0.00 14 13:03 0.07 21 13:10 0.00

[0078] The calibration standard value of the mine is 1.5%. The records where the value obtained from the table is greater than 0.4K, i.e., 0.6, are (12:57, 0.67), (12:58, 0.67), and (13:01, 1.01). Then, the calibration interval is obtained by scaling 5 minutes before and after [12:57, 13:01], which is [12:52, 13:06]. Subsequently, the values in this area are filled based on the interpolation algorithm.

[0079] The mine adopts a three-shift system with a shift cycle of 480 minutes. The designed cyclic footage is 0.8m, and the single-cycle operation period is 150 minutes. During this period, the working face advanced 215.7m, with an average daily advance of 2.37m. The distance between the return air and the working face sensor increased from 74m to 289.7m. The average wind speed of the working face is 0.56m / s. Then, the minimum sensor delay is 2.2 minutes, and the maximum is 8.6 minutes. The maximum of the supremum of the delay, which is 9 minutes, is taken as the reference delay.

[0080] As Figure 8 、 Figure 9 shown, for the data of the working face gas sensor A1 and the return air gas sensor A2, first, there are 2961 and 1654 after identification based on the 3σ rule of the difference value; then, the abnormal points where the difference value is greater than the moving average are filtered out, which are 31 and 10 respectively. The first ten abnormal value points of both are shown in Table 4.

[0081] Table 4

[0082]

[0083] By traversing the pairs of abnormal time points of the two sensors in a loop, it is determined that all time points are locally abnormal. After deletion, the adjacent values are used for filling.

[0084] Finally, 131,040 four-item data are standardized to form a basic dataset, which is divided into a training set of 91,712 items, a test set of 26,176 items, and a validation set of 13,120 items (all are integer multiples of 64) according to the ratio of 7:2:1. The basic parameters of the model are set as shown in Table 5.

[0085] Table 5

[0086] Parameter Name Parameter Value Parameter Name Parameter Value Number of Iterations 10 Learning Rate 0.001 Batch Size 32 Length of Look-back Window 60 Length of Prior Sequence 30 Length of Prediction Window 60 Number of encoder Layers 2 Number of decoder Layers 1 Number of Neurons in FFN Layer 32 Number of Neurons in Linear Layer 16 Number of Top Frequencies 5 Shift Cycle 480 Cycle of Cyclic Operation 150 γ Value of Loss Function 0.8

[0087] The Soft-DTW loss function is used for model training, with the hyperparameter γ = 0.8. The alignment matrix weight is dynamically adjusted through a differentiable minimization function to optimize the temporal alignment between the predicted sequence and the actual sequence.

[0088] The Adam optimizer is used for model training, with the initial learning rate set to 0.001, and the learning rate is dynamically adjusted according to the validation set loss during the training process, with the range controlled between 10 -5 ~10 -3 .

[0089] The model is trained using a single NVIDIA A100 graphics card, with 6 effective iterative trainings. The total training and testing time is about 40 minutes. The mean value of the loss function of the training set drops from -68.30 to -73.44, and the mean value of the loss function of the validation set drops from -72.79 to -75.92. Finally, the mean value of the loss function of the test set is -74.62. As Figure 10 shown, the fluctuation laws of the actual value and the predicted value curves after standardization in this training are highly consistent.

[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 restrictive. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the present technical solution, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A method for long-term prediction of gas concentration in the working face based on the DTW-BiT-TimesNet model, characterized in that: Including the following steps: S1: Collect the monitoring data of the gas sensors at the working face, the gas sensors at the return air location, the wind speed sensors, and the dust sensors in the coal mine safety monitoring system for no less than 3 months under continuous and stable mining conditions, extract the sub-averages and record them as data column A i , i ∈ [1, 4]; S2: Conduct verification data filtering and filling on data columns A1 and A2, and perform anomaly identification and correction based on time lag and correlation; perform data standardization on data columns A1, A2, A3, and A4, generate a basic data set, and divide it into a training set, a test set, and a validation set according to the ratio of 7:2:1; S3: Using A1, A2, A3, and A4 as input variables, with A1 as the predicted variable, based on the mine shift cycle T bc and the cyclic operation cycle T xh Construct a DTW-BiT-TimesNe model, and use the gradient descent method and the Soft-DTW loss function to train and validate the model, and output the long-term prediction result of the gas concentration A1 in the working face.

2. The method for long-term prediction of working face gas concentration based on the DTW-BiT-TimesNet model according to claim 1, characterized in that: The verification data filtering includes: Identify abnormal data according to the regular calibration rules of the mine, set the calibration time interval [t a , t b , and the calibration value K. Traverse this interval to calculate the absolute value of the difference v j , j ∈ N + . When two consecutive difference values are greater than the specified threshold and the monitored value exceeds 0.4K, mark it as an abnormal point, and extend the abnormal point interval to the left and right safety thresholds m. Finally, fill the data in the abnormal interval through the nearest interpolation algorithm.

3. The method for long-term prediction of working face gas concentration based on the DTW-BiT-TimesNet model according to claim 2, wherein: The anomaly identification and correction based on time lag and correlation includes: The first-order difference operation is performed on A1 and A2 to obtain B1 and B2. The moving average with a window of w is performed on A1 and A2, and then shifted forward by w / 2 bits to complete filling, resulting in C1 and C2, where w is an even number. Based on the 3σ statistical anomaly recognition method, after identifying the differential values at the abnormal time points in B1 and B2, the parts with differential values greater than C1 and C2 are selected and determined as the abnormal pending time points T c1 and T c2 . Combining the time delay Δt between the working face and the return air sensor, it is judged whether the abnormal points are paired. If the abnormal points are paired, they are determined as real anomalies and retained; otherwise, they are determined as local anomalies, deleted, and then interpolated and filled 4. The method for long-term prediction of working face gas concentration based on the DTW-BiT-TimesNet model according to claim 1, wherein: The construction of the DTW-BiT-TimesNe model includes: Perform frequency domain analysis on the input data and select the period T corresponding to the k values with the highest frequencies i , where i ∈ [1, k], and the mine shift period T bc , the cyclic operation period T xh Divide the data into k + 2 two-dimensional data with the length of the period value; after convolving and extracting features from the k + 2 two-dimensional arrays based on the Inception Block in TimesNet and then converting them back to k + 2 one-dimensional arrays, analyze the temporal features layer by layer through the following weighted summation formula: where l is the number of layers of the time series analysis block, is the output of the (l - 1)-th layer time series analysis block and the input of the l-th layer time series analysis block; i ∈ [1, k + 2] is the data after the (l - 1)-th layer time series analysis block analyzes and converts it into one dimension; Concat(·) is the merging of one-dimensional arrays; Softmax(·) is the normalized exponential function; i ∈ [1, k] is the amplitude corresponding to the Top K frequencies after the (l - 1)-th layer frequency domain decomposition; w bc , w xh are the undetermined weight parameters corresponding to the decomposition of the mine shift cycle and the cyclic operation cycle respectively.

5. The method for long-term prediction of working face gas concentration based on the DTW-BiT-TimesNet model according to claim 1, characterized in that: The Soft-DTW loss function is defined as: where y * is the actual value, is the predicted value, is the initialized cost matrix, solved with reference to the DTW algorithm; is the inner product of the alignment matrix A and the cost matrix; min γ {·} is a differentiable minimization function containing the hyperparameter γ, defined as: where a i , i ∈ [1, n] is the term for which the minimum value needs to be found, and n is the number of terms.

6. The method for long-term prediction of working face gas concentration based on the DTW-BiT-TimesNet model according to claim 1, characterized in that: In S2, the data standardization includes: performing standardization processing on A1, A2, A3, and A4, and generating a basic data set with a unified scale after eliminating the dimension difference.

7. The method for long-term prediction of working face gas concentration based on the DTW-BiT-TimesNet model according to claim 2, characterized in that: The length of the verification time interval is one day, i.e., t b -t a = 1440 min, and the value of the left and right safety thresholds m is 5 min.

8. The method for long-term prediction of working face gas concentration based on the DTW-BiT-TimesNet model according to claim 3, wherein: The value range of the hyperparameter γ of the Soft-DTW loss function is 0 to 1.

9. The method for long-term prediction of working face gas concentration based on the DTW-BiT-TimesNet model according to claim 4, characterized in that: The selection criterion for the Top K frequencies is the first 5 frequencies arranged in descending order of amplitude.

10. The method for long-term prediction of working face gas concentration based on the DTW-BiT-TimesNet model according to claim 1, characterized in that: During the model training, the dynamic adjustment range of the learning rate of the gradient descent method is 10 -5 ~10 -3 .

Images