Gas concentration prediction method for tunneling working face in coal mine underground

Abstract

The present application relates to a coal mine underground tunneling working face gas concentration prediction method, belonging to the technical field of coal mine intelligent prediction. The method comprises: obtaining the original time series data sequence collected by multiple types of sensors in the preset time period in the coal mine underground tunneling working face, and performing standardization processing to obtain the standard time series data sequence corresponding to each type of sensor, then using multilayer perception to perform linear projection to obtain the real-time high-dimensional vector corresponding to each type of sensor; constructing and training a gas concentration prediction model; inputting the real-time high-dimensional vector corresponding to each type of sensor into the gas concentration prediction model, and outputting the gas concentration prediction function in the preset prediction time length by the gas concentration prediction model; discretizing the gas concentration prediction function according to the self-defined prediction time interval to obtain the gas concentration prediction sequence in the preset prediction time length. The present application can improve the accuracy and prediction efficiency of the prediction result.

Description

Technical Field

[0001] This invention relates to the field of intelligent prediction technology in coal mines, and in particular to a method for predicting gas concentration in underground tunneling faces of coal mines. Background Technology

[0002] Abnormal gas accumulation in underground coal mine tunneling faces is a significant factor in inducing gas explosions and outbursts. Developing an effective method for predicting gas concentration in underground coal mine tunneling faces is of great importance for improving the ability to predict and warn of gas disasters in tunneling faces and for realizing intelligent tunneling.

[0003] Currently, methods for predicting methane concentration at tunneling faces mainly employ time series models or simple machine learning models. However, these methods typically rely solely on methane concentration data collected by methane concentration sensors, neglecting the influence of other data on methane concentration. This results in low accuracy and significant prediction bias. Furthermore, since the data acquisition frequency of each sensor at the tunneling face is dynamically adjusted, any change in the acquisition frequency of any sensor necessitates retraining the methane concentration prediction model, leading to cumbersome operations and low prediction efficiency. Summary of the Invention

[0004] To solve the above-mentioned technical problems, this invention provides a method for predicting gas concentration in underground coal mine tunneling faces. The technical solution of this invention is as follows: Methods for predicting methane concentration in underground coal mine tunneling faces include: S1, acquire the original time-series data sequence within a preset time period collected in real time by multiple types of sensors in the underground tunneling face of the coal mine. The multiple types of sensors include gas concentration sensors directly coupled to gas concentration and other types of sensors indirectly coupled to gas concentration. S2, standardize the raw time-series data sequences collected by each type of sensor to obtain the standard time-series data sequences corresponding to each type of sensor; S3, the standard time series data sequence corresponding to each type of sensor is linearly projected using a multilayer perceptron to obtain the real-time high-dimensional vector corresponding to each type of sensor; S4, Construct and train a gas concentration prediction model; S5, input the real-time high-dimensional vector corresponding to each type of sensor into the gas concentration prediction model, and the gas concentration prediction model outputs the gas concentration prediction function within the preset prediction time. S6. Discretize the gas concentration prediction function according to the custom prediction time interval to obtain the gas concentration prediction sequence within the preset prediction time.

[0005] Preferably, S2 includes: S21, determine the periodicity of the raw time-series data sequence acquired by each type of sensor; S22, Divide the original time-series data sequence collected by each type of sensor into multiple data blocks according to the periodicity, and obtain multiple data blocks corresponding to each type of sensor; S23, a sliding window is used to slide across each data block corresponding to each type of sensor, the local mean and local standard deviation of the original time series data in each sliding window are calculated, and the last original time series data in each sliding window is standardized based on the local mean and local standard deviation to obtain the standard time series data sequence corresponding to each type of sensor.

[0006] Preferably, S4 includes: S41, acquire historical time-series data sequences collected by multiple types of sensors, and process the historical time-series data sequences collected by each type of sensor to obtain the historical high-dimensional vector corresponding to each type of sensor. The historical high-dimensional vector corresponding to each type of sensor is composed of the historical mapping values ​​of a first preset number of historical time nodes. S42, construct the state-space model corresponding to each type of sensor; S43, create an explicit state dependency matrix and a time mask matrix based on the historical time series data sequences collected by each type of sensor, and determine the weighted fusion mask matrix based on the explicit state dependency matrix and the time mask matrix; S44, Discretize and align the sampling frequency of the state space model corresponding to each type of sensor to obtain the discretized state vector corresponding to each type of sensor; S45. Based on the state space model and its discretized state vector, weighted fusion mask matrix and historical high-dimensional vector corresponding to each type of sensor, construct the historical gas concentration feature vector for each historical time node. S46, using a multilayer perceptron to map the historical gas concentration feature vectors of all historical time points into a second preset number of prediction coefficients; S47, construct a gas concentration prediction function based on a second preset number of prediction coefficients, a preset basis function group, and a preset prediction duration, wherein the number of preset basis functions in the preset basis function group is the second preset number; S48, the loss function for constructing the gas concentration prediction model; S49. Initialize the learnable parameters of the weighted fusion mask matrix, the learnable parameters of the state space model corresponding to each type of sensor, the learnable parameters of the historical gas concentration feature vector at each historical time point, and the learnable parameters of the multilayer perceptron. Calculate the initial historical gas concentration prediction value based on the initialization results and the gas concentration prediction function. Iteratively update the learnable parameters of the weighted fusion mask matrix, the learnable parameters of the state space model corresponding to each type of sensor, the learnable parameters of the historical gas concentration feature vector at each historical time point, and the learnable parameters of the multilayer perceptron through backpropagation algorithm based on the error and loss function between the initial historical gas concentration prediction value and the actual historical gas concentration value. Stop training when the convergence condition is met, and obtain the trained gas concentration prediction model.

[0007] Preferably, the state-space model corresponding to the s-th type of sensor constructed by S42 is expressed as formula (1): (1); In formula (1), This represents the state space of the t-th historical time node corresponding to the s-th type of sensor. This represents the state transition matrix corresponding to the s-th type of sensor. This represents the state space of the (t-1)th historical time node corresponding to the s-th type of sensor. This represents the input influence matrix corresponding to the s-th type of sensor. Let represent the historical mapping value at the (t-1)th historical time node in the historical high-dimensional vector corresponding to the s-th type of sensor. The learnable parameters of the state-space model corresponding to the s-th type of sensor include... , and , Let represent the state space of the initial historical time node corresponding to the s-th type of sensor.

[0008] Preferably, S43 includes: S431, Construct an explicit state dependency matrix based on the dependency relationship between the historical gas concentration time series data sequence collected by the gas concentration sensor and the historical time series data sequence collected by each other type of sensor; S432, create a time mask matrix based on the temporal sequence of each historical time series data in the historical time series data sequence collected by each type of sensor; S433: The explicit state dependency matrix and the time mask matrix are decomposed into low-rank matrices using matrix factorization, and the decomposition results are weighted and fused to obtain a weighted fused mask matrix.

[0009] Preferably, step S433 involves weighted fusion of the decomposition results to obtain a weighted fusion mask matrix. When, this is achieved through formula (2): (2); In formula (2), and All of these are learnable weight coefficients. Represents the time mask matrix The result of low-rank decomposition. Represents the explicit state dependency matrix The result of low-rank decomposition, where T denotes the transpose of the matrix. Indicates the number based on sensor type The generated diagonal matrix, Indicates the number based on historical time points. The generated diagonal matrix, It represents the Kronecker product.

[0010] Preferably, S45 includes: S451, combine the discretized state vectors corresponding to all types of sensors to obtain the discretized state matrix; S452, combine the historical high-dimensional vectors corresponding to all types of sensors to obtain the historical high-dimensional matrix; S453, based on the state space model, discretized state matrix, weighted fusion mask matrix and historical high-dimensional matrix corresponding to the gas concentration sensor, construct the historical gas concentration feature vector for each historical time node.

[0011] Preferably, step S453 is based on the state-space model corresponding to the gas concentration sensor. Discretized state matrix Weighted fusion mask matrix and historical high-dimensional matrix The historical gas concentration feature vector at the t-th historical time node is constructed using formula (3). : (3); In formula (3), Represents the observation matrix. This indicates the output influence matrix. This represents a vector concatenation operation. The attention pooling operation represents the learnable parameters of the historical gas concentration feature vector at the t-th historical time node, including: and .

[0012] Preferably, in step S46, a multilayer perceptron is used to generate the historical gas concentration feature vector at the t-th historical time point. Mapped to prediction coefficients When, this is achieved through formula (4): (4); In formula (4), Describes the multilayer perceptron function. This represents the gated residual activation network function. This represents the learnable weight matrix. This represents a learnable bias vector. This represents the learnable residual weight parameters. The value of n is the second preset quantity.

[0013] All of the above-mentioned optional technical solutions can be combined arbitrarily, and the present invention will not provide a detailed description of the structure after each combination.

[0014] The beneficial effects of this invention, achieved through the above scheme, are as follows: By acquiring the original time-series data sequences within a preset time period collected by gas concentration sensors directly coupled to gas concentration and other types of sensors indirectly coupled to gas concentration in the tunneling face, the method provided by this invention considers the influence of multiple data types simultaneously when predicting gas concentration, resulting in more accurate prediction results. By standardizing the original time-series data sequences collected by each type of sensor, a standard time-series data sequence corresponding to each type of sensor is obtained. Then, the standard time-series data sequences corresponding to each type of sensor are linearly projected to obtain the real-time high-dimensional vector corresponding to each type of sensor. A gas concentration prediction model is then constructed and trained, and the real-time high-dimensional vector corresponding to each type of sensor is input into the gas concentration prediction model. The gas concentration prediction model outputs a gas concentration prediction function within a preset prediction time period. This ensures that the method provided by this invention is independent of the sampling frequency of each type of sensor. When the sampling frequency of any type of sensor changes, it is not necessary to retrain the model, thereby improving prediction efficiency.

[0015] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method for predicting gas concentration in underground tunneling faces in coal mines provided by the present invention. Detailed Implementation

[0017] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0018] like Figure 1 As shown in the embodiment of the present invention, the method for predicting gas concentration in underground coal mine tunneling faces includes: S1, acquire the original time-series data sequence within a preset time period collected in real time by multiple types of sensors in the underground tunneling face of the coal mine. The multiple types of sensors include gas concentration sensors directly coupled to gas concentration and other types of sensors indirectly coupled to gas concentration.

[0019] Specifically, the types of sensors include gas concentration sensors, wind speed sensors, current sensors, and temperature sensors. The preset time period in this embodiment is generally 30 minutes. In actual tunneling, it was found that gas concentration increases when the tunneling machine cuts and decreases when wind speed increases. Therefore, the gas concentration at the tunneling face is related not only to the gas concentration data collected by the gas concentration sensor but also to the data collected by other types of sensors. Therefore, the method provided in this embodiment of the invention, when predicting gas concentration, also comprehensively considers other types of sensors indirectly coupled to gas concentration. These other types of sensors indirectly coupled to gas concentration include wind speed sensors, current sensors, and temperature sensors. The types of original time-series data sequences include gas concentration time-series data sequences, wind speed time-series data sequences, current time-series data sequences, and temperature time-series data sequences. Since the sampling frequency of each type of sensor is different, this embodiment of the invention does not resample the original time-series data sequences collected by each type of sensor; instead, it directly retains the original timestamp and sampling interval information of the original time-series data sequences collected by each type of sensor.

[0020] S2, standardizes the raw time-series data sequences collected by each type of sensor to obtain the standard time-series data sequences corresponding to each type of sensor.

[0021] Specifically, since the sampling frequencies of various sensors are inconsistent, they cannot be uniformly standardized. To avoid information leakage, this embodiment of the invention employs local masking standardization. When standardizing a specific original time series data in a given original time series data sequence using local masking, only the original time series data corresponding to the time node and the original time series data preceding it can be used; no future information (original time series data following the time node corresponding to the original time series data) is used.

[0022] In one specific embodiment, S2 includes: S21, determine the periodicity of the raw time-series data sequence acquired by each type of sensor.

[0023] Specifically, by analyzing the variation patterns of the raw time-series data sequences collected by each type of sensor, the periodicity of the raw time-series data sequences collected by each type of sensor can be obtained.

[0024] S22, the original time-series data sequence collected by each type of sensor is divided into multiple data blocks according to the periodicity, so as to obtain multiple data blocks corresponding to each type of sensor.

[0025] Specifically, the number of data blocks corresponding to each type of sensor is the same as the periodicity of the original time-series data sequence acquired by each type of sensor. For example, if the original time-series data sequence acquired by a certain type of sensor has four periods, then the original time-series data sequence acquired by that type of sensor is divided into four data blocks.

[0026] S23, a sliding window is used to slide across each data block corresponding to each type of sensor, the local mean and local standard deviation of the original time series data in each sliding window are calculated, and the last original time series data in each sliding window is standardized based on the local mean and local standard deviation to obtain the standard time series data sequence corresponding to each type of sensor.

[0027] Specifically, in this embodiment of the invention, the size of the sliding window is 32 or 64, and the sliding step is 1. When standardizing the original time-series data sequence corresponding to a certain type of sensor, only the last original time-series data in the sliding window is standardized each time until the end of the original time-series data sequence is reached, completing the standardization process. Assuming the sliding window size is 32, when standardizing the last original time-series data in a certain sliding window, the local mean and local variance of the 32 original time-series data in the sliding window are first calculated. Then, the local mean and local variance are used to standardize the last original time-series data in the sliding window using the mean-variance standardization algorithm to obtain the standard time-series data of the last original time-series data in the sliding window. It should be noted that, since this embodiment of the invention only performs standardization processing on the last original time-series data in each sliding window, the first 31 original time-series data are not standardized when standardizing the original time-series data sequence corresponding to this type of sensor. By combining the standard time series data of all raw time series data except the first 31 raw time series data points corresponding to this type of sensor, a standard time series data sequence corresponding to this type of sensor is obtained. Through local normalization, the accuracy of the original time series data is preserved.

[0028] S3. The standard time-series data sequence corresponding to each type of sensor is linearly projected using a multilayer perceptron to obtain the real-time high-dimensional vector corresponding to each type of sensor.

[0029] Specifically, the standard time-series data sequence corresponding to each type of sensor is input into a multilayer perceptron (MLP), and the MLP outputs a real-time high-dimensional vector corresponding to each type of sensor. The real-time high-dimensional vectors for all types of sensors have the same dimension, which is a first preset number of dimensions, for example, 128 dimensions. Each type of sensor's real-time high-dimensional vector consists of real-time mapping values ​​for the first preset number of dimensions.

[0030] S4. Construct and train a gas concentration prediction model.

[0031] The gas concentration prediction model is used to predict the gas concentration prediction function within a preset prediction time period.

[0032] In one specific embodiment, S4 includes: S41, acquire historical time-series data sequences collected by multiple types of sensors, and process the historical time-series data sequences collected by each type of sensor to obtain the historical high-dimensional vector corresponding to each type of sensor. The historical high-dimensional vector corresponding to each type of sensor is composed of the historical mapping values ​​of a first preset number of historical time nodes.

[0033] Specifically, the principle of processing the historical time-series data sequences collected by each type of sensor to obtain the historical high-dimensional vector corresponding to each type of sensor is the same as the principle of obtaining the corresponding real-time high-dimensional vector based on the original time-series data sequences collected by each type of sensor, and will not be elaborated here.

[0034] S42, construct the state-space model corresponding to each type of sensor.

[0035] Specifically, when constructing a state space model for a certain type of sensor, it is necessary to randomly initialize all parameters of the state space model, then map the historical high-dimensional vectors of that type of sensor into the state space, and generate new state representations through iterative updates until the iteration is complete to obtain the state space model for that type of sensor.

[0036] In a specific embodiment, the state space model corresponding to the s-th type of sensor constructed by S42 is expressed as formula (1): (1); In formula (1), This represents the state space of the t-th historical time node corresponding to the s-th type of sensor. This represents the state transition matrix corresponding to the s-th type of sensor. This represents the state space of the (t-1)th historical time node corresponding to the s-th type of sensor. This represents the input influence matrix corresponding to the s-th type of sensor. Let represent the historical mapping value at the (t-1)th historical time node in the historical high-dimensional vector corresponding to the s-th type of sensor. The learnable parameters of the state-space model corresponding to the s-th type of sensor include... , and , Let represent the state space of the initial historical time node (the 0th historical time node) corresponding to the s-th type of sensor.

[0037] S43, create an explicit state dependency matrix and a time mask matrix based on the historical time-series data sequences collected by each type of sensor, and determine a weighted fusion mask matrix based on the explicit state dependency matrix and the time mask matrix.

[0038] Among them, the explicit state dependency matrix represents the physical dependency between the historical gas concentration time series data collected by the gas concentration sensor and the historical time series data collected by other types of sensors, and the time mask matrix represents the temporal order of each historical time series data in the historical time series data sequence.

[0039] Specifically, the explicit state dependency matrix is ​​constructed based on the physical dependencies between historical gas concentration time-series data and historical time-series data collected by other types of sensors. The initial values ​​of the elements in the explicit state dependency matrix include 1 and 0. 1 indicates a physical dependency between the historical gas concentration time-series data and historical time-series data collected by another type of sensor, while 0 indicates no physical dependency between the two historical time-series data collected by other types of sensors. The existence of physical dependencies is determined through empirical review. For example, if the gas concentration increases when the tunnel boring machine cuts (increases current) and decreases when the wind speed increases, it indicates that both the historical current time-series data collected by the current sensor and the historical wind speed time-series data collected by the wind speed sensor have physical dependencies on the historical gas concentration time-series data. In this case, the elements representing the physical dependency between the historical current time-series data and the historical gas concentration time-series data are initialized to 1 in the explicit state dependency matrix, and the elements representing the physical dependency between the historical wind speed time-series data and the historical gas concentration time-series data are also initialized to 1. For example, if there is no physical dependency between the historical vibration time-series data sequence collected by the vibration sensor and the historical gas concentration time-series data sequence, then the element representing the physical dependency between the two sequences in the explicit state dependency matrix will be set to 0. It should be noted that during subsequent model training, the elements initialized to 1 in the explicit state dependency matrix need to be iteratively updated in real time according to their coupling degree with the historical gas concentration time-series data sequence; some will increase, some will decrease, until the iteration ends.

[0040] If the acquisition time of one historical time series data point is earlier than the acquisition time of another historical time series data point, then the element in the time mask matrix representing the temporal relationship between the two historical time series data points is 1; otherwise, it is 0. Therefore, the time mask matrix is ​​a lower triangular matrix, with all parameters of the lower triangular matrix being 1 and all parameters of the upper triangular matrix being 0.

[0041] Based on the above, in a specific embodiment, S43 includes: S431, construct an explicit state dependency matrix based on the dependency relationship between the historical gas concentration time series data sequence collected by the gas concentration sensor and the historical time series data sequence collected by each other type of sensor.

[0042] S432, create a time mask matrix based on the temporal sequence of historical time-series data collected by each type of sensor.

[0043] Specifically, when constructing the gas concentration prediction model, the temporal sequence of the original time series data for each historical time node is constrained by considering the time mask matrix. This ensures that when training the gas concentration prediction model, only the historical time series data before each historical time node is referenced, thus avoiding premature glimpses of future information during the prediction process.

[0044] S433: The explicit state dependency matrix and the time mask matrix are decomposed into low-rank matrices using matrix factorization, and the decomposition results are weighted and fused to obtain a weighted fused mask matrix.

[0045] In a specific embodiment, step S433 performs weighted fusion on the decomposition results to obtain a weighted fusion mask matrix. When, this is achieved through formula (2): (2); In formula (2), and All of these are learnable weight coefficients. Represents the time mask matrix The result of low-rank decomposition. Represents the explicit state dependency matrix The result of low-rank decomposition, where T denotes the transpose of the matrix. Indicates the number based on sensor type The generated diagonal matrix, Indicates the number of historical time points The generated diagonal matrix, This represents the Kronecker product. The learnable parameters of the weighted fusion mask matrix include... and .

[0046] By generating a weighted fusion mask matrix, the method provided in this embodiment of the invention can fully extract the hidden states and nonlinear dependencies between variables in the historical time series data sequences corresponding to various types of sensors and the historical gas concentration time series data sequences when training the gas concentration prediction model. At the same time, it avoids learning invalid dependencies between variables, thereby reducing the computational complexity of the gas concentration prediction model and improving the prediction efficiency of the gas concentration prediction model.

[0047] S44. Discretize and align the sampling frequency of the state space model corresponding to each type of sensor to obtain the discretized state vector corresponding to each type of sensor.

[0048] Specifically, since the state-space model is continuous, subsequent predictions require discretizing the state-space model for each type of sensor and aligning its sampling frequencies. The formula for discretizing and aligning the sampling frequencies of the state-space model for a specific type of sensor is as follows: ;in, This represents the relative scaling factor between the sampling frequency of this type of sensor and the sampling frequency of the gas concentration sensor. It is calculated as the ratio of the sampling frequency of this type of sensor to the sampling frequency of the gas concentration sensor. The discretized state vectors for all types of sensors contain the same number of elements.

[0049] S45. Based on the state space model and its discretized state vector, weighted fusion mask matrix and historical high-dimensional vector corresponding to each type of sensor, construct the historical gas concentration feature vector for each historical time node.

[0050] In one specific embodiment, S45 includes: S451, combine the discretized state vectors corresponding to all types of sensors to obtain the discretized state matrix.

[0051] Specifically, one column in the discretized state matrix is ​​the discretized state vector corresponding to a sensor type, and the number of rows in the discretized state matrix is ​​the same as the number of elements in the discretized state vector.

[0052] S452, combine the historical high-dimensional vectors corresponding to all types of sensors to obtain the historical high-dimensional matrix.

[0053] Specifically, one column in the historical high-dimensional matrix is ​​a historical high-dimensional vector corresponding to a certain type of sensor, and the number of rows in the historical high-dimensional matrix is ​​the same as the dimension of the historical high-dimensional vector.

[0054] S453, based on the state space model, discretized state matrix, weighted fusion mask matrix and historical high-dimensional matrix corresponding to the gas concentration sensor, construct the historical gas concentration feature vector for each historical time node.

[0055] In one specific embodiment, S453 is based on the state-space model corresponding to the gas concentration sensor. Discretized state matrix Weighted fusion mask matrix and historical high-dimensional matrix The historical gas concentration feature vector at the t-th historical time node is constructed using formula (3). : (3); In formula (3), Represents the observation matrix. This indicates the output influence matrix. This represents a vector concatenation operation. The attention pooling operation represents the learnable parameters of the historical gas concentration feature vector at the t-th historical time node, including: and .

[0056] S46 uses a multilayer perceptron to map the historical gas concentration feature vectors of all historical time points into a second preset number of prediction coefficients.

[0057] Specifically, a multilayer perceptron is used to transform all historical gas concentration feature vectors into a coefficient space to obtain a second preset number of prediction coefficients.

[0058] In a specific embodiment, step S46 uses a multilayer perceptron to process the historical gas concentration feature vector at the t-th historical time point. Mapped to prediction coefficients When, this is achieved through formula (4): (4); In formula (4), Describes the multilayer perceptron function. This represents the gated residual activation network function. This represents the learnable weight matrix. This represents a learnable bias vector. This represents the learnable residual weight parameters. The value of n is a second preset number. The learnable parameters of a multilayer perceptron include... , and .

[0059] Specifically, This approach allows the output prediction coefficients to possess both nonlinear gating characteristics and retain the loss information of the original state, thereby enhancing the nonlinear features.

[0060] S47, construct a gas concentration prediction function based on a second preset number of prediction coefficients, a preset basis function group, and a preset prediction duration, wherein the number of preset basis functions in the preset basis function group is the second preset number.

[0061] Specifically, the number of preset basis functions and the form of the preset basis function set were determined through expert evaluation. Gas concentration prediction function. It can be represented as: ;in, This represents the i-th prediction coefficient. This represents the i-th preset basis function in the preset basis function set. Indicates time parameter, These represent the start and end times of the preset prediction duration, respectively.

[0062] S48, the loss function for constructing the gas concentration prediction model.

[0063] Specifically, in this embodiment of the invention, the loss function of the gas concentration prediction model introduces physical constraint loss and time constraint loss on the basis of mean square error. The specific formula is as follows: ;in, Indicates loss, This indicates the error between the predicted historical gas concentration and the actual historical gas concentration. Indicates physical loss. Indicates time loss, and These are adjustable parameters, where the loss is adjusted when the physical constraint loss and temporal constraint loss meet the preset physical and temporal dependencies during training. Reduce, or conversely, lose Then it increases.

[0064] S49. Initialize the learnable parameters of the weighted fusion mask matrix, the learnable parameters of the state space model corresponding to each type of sensor, the learnable parameters of the historical gas concentration feature vector at each historical time point, and the learnable parameters of the multilayer perceptron. Calculate the initial historical gas concentration prediction value based on the initialization results and the gas concentration prediction function. Iteratively update the learnable parameters of the weighted fusion mask matrix, the learnable parameters of the state space model corresponding to each type of sensor, the learnable parameters of the historical gas concentration feature vector at each historical time point, and the learnable parameters of the multilayer perceptron through backpropagation algorithm based on the error and loss function between the initial historical gas concentration prediction value and the actual historical gas concentration value. Stop training when the convergence condition is met, and obtain the trained gas concentration prediction model.

[0065] The convergence condition can be reaching a preset number of iterations or the loss being less than a preset loss threshold. For the principle of iterative updates using the backpropagation algorithm, please refer to the iterative update principles of existing neural network models; this embodiment of the invention will not elaborate on this in detail.

[0066] S5 inputs the real-time high-dimensional vector corresponding to each type of sensor into the gas concentration prediction model, and the gas concentration prediction model outputs the gas concentration prediction function within the preset prediction time.

[0067] S6. Discretize the gas concentration prediction function according to the custom prediction time interval to obtain the gas concentration prediction sequence within the preset prediction time.

[0068] Specifically, the custom prediction time interval can be adaptively adjusted based on the sensor's sampling frequency. The specific adjustment involves customizing the prediction time interval. It can be adjusted to: ;in, This represents the original prediction time interval, which is the prediction time interval before the sensor sampling frequency changes. The relative scaling factor represents the sensor sampling frequency, and is calculated as the ratio of the original sampling frequency to the new sampling frequency. The method provided in this invention can automatically adapt to discrete prediction outputs at different sampling frequencies.

[0069] Since the method provided in this embodiment of the invention uses raw time-series data sequences collected in real time by multiple types of sensors within a preset time period when triggering gas concentration prediction, it is not necessary to retrain the gas concentration prediction model even if the sampling frequency of a certain type of sensor changes, thus improving the efficiency of gas concentration prediction.

[0070] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting gas concentration in underground coal mine tunneling faces, characterized in that, include: S1, acquire the original time-series data sequence within a preset time period collected in real time by multiple types of sensors in the underground tunneling face of the coal mine. The multiple types of sensors include gas concentration sensors directly coupled to gas concentration and other types of sensors indirectly coupled to gas concentration. S2, standardize the raw time-series data sequences collected by each type of sensor to obtain the standard time-series data sequences corresponding to each type of sensor; Wherein, S2 includes: S21, determine the periodicity of the raw time-series data sequence acquired by each type of sensor; S22, Divide the original time-series data sequence collected by each type of sensor into multiple data blocks according to the periodicity, and obtain multiple data blocks corresponding to each type of sensor; S23, a sliding window is used to slide over each data block corresponding to each type of sensor, the local mean and local standard deviation of the original time series data in each sliding window are calculated, and the last original time series data in each sliding window is standardized based on the local mean and local standard deviation to obtain the standard time series data sequence corresponding to each type of sensor; S3, the standard time series data sequence corresponding to each type of sensor is linearly projected using a multilayer perceptron to obtain the real-time high-dimensional vector corresponding to each type of sensor; S4, Construct and train a gas concentration prediction model; S5, input the real-time high-dimensional vector corresponding to each type of sensor into the gas concentration prediction model, and the gas concentration prediction model outputs the gas concentration prediction function within the preset prediction time. S6. Discretize the gas concentration prediction function according to the custom prediction time interval to obtain the gas concentration prediction sequence within the preset prediction time.

2. The method for predicting gas concentration in underground coal mine tunneling faces according to claim 1, characterized in that, S4 includes: S41, acquire historical time-series data sequences collected by multiple types of sensors, and process the historical time-series data sequences collected by each type of sensor to obtain the historical high-dimensional vector corresponding to each type of sensor. The historical high-dimensional vector corresponding to each type of sensor is composed of the historical mapping values ​​of a first preset number of historical time nodes. S42, construct the state-space model corresponding to each type of sensor; S43, create an explicit state dependency matrix and a time mask matrix based on the historical time series data sequences collected by each type of sensor, and determine the weighted fusion mask matrix based on the explicit state dependency matrix and the time mask matrix; S44, Discretize and align the sampling frequency of the state space model corresponding to each type of sensor to obtain the discretized state vector corresponding to each type of sensor; S45. Based on the state space model and its discretized state vector, weighted fusion mask matrix and historical high-dimensional vector corresponding to each type of sensor, construct the historical gas concentration feature vector for each historical time node. S46, using a multilayer perceptron to map the historical gas concentration feature vectors of all historical time points into a second preset number of prediction coefficients; S47, construct a gas concentration prediction function based on a second preset number of prediction coefficients, a preset basis function group, and a preset prediction duration, wherein the number of preset basis functions in the preset basis function group is the second preset number; S48, the loss function for constructing the gas concentration prediction model; S49. Initialize the learnable parameters of the weighted fusion mask matrix, the learnable parameters of the state space model corresponding to each type of sensor, the learnable parameters of the historical gas concentration feature vector at each historical time point, and the learnable parameters of the multilayer perceptron. Calculate the initial historical gas concentration prediction value based on the initialization results and the gas concentration prediction function. Iteratively update the learnable parameters of the weighted fusion mask matrix, the learnable parameters of the state space model corresponding to each type of sensor, the learnable parameters of the historical gas concentration feature vector at each historical time point, and the learnable parameters of the multilayer perceptron through backpropagation algorithm based on the error and loss function between the initial historical gas concentration prediction value and the actual historical gas concentration value. Stop training when the convergence condition is met, and obtain the trained gas concentration prediction model.

3. The method for predicting gas concentration in underground coal mine tunneling faces according to claim 2, characterized in that, The state-space model corresponding to the s-th type of sensor constructed by S42 is expressed by formula (1): （1）； In formula (1), This represents the state space of the t-th historical time node corresponding to the s-th type of sensor. This represents the state transition matrix corresponding to the s-th type of sensor. This represents the state space of the (t-1)th historical time node corresponding to the s-th type of sensor. This represents the input influence matrix corresponding to the s-th type of sensor. Let represent the historical mapping value at the (t-1)th historical time node in the historical high-dimensional vector corresponding to the s-th type of sensor. The learnable parameters of the state-space model corresponding to the s-th type of sensor include... , and , Let represent the state space of the initial historical time node corresponding to the s-th type of sensor.

4. The method for predicting gas concentration in underground coal mine tunneling faces according to claim 2, characterized in that, S43 includes: S431, Construct an explicit state dependency matrix based on the dependency relationship between the historical gas concentration time series data sequence collected by the gas concentration sensor and the historical time series data sequence collected by each other type of sensor; S432, create a time mask matrix based on the temporal sequence of each historical time series data in the historical time series data sequence collected by each type of sensor; S433: The explicit state dependency matrix and the time mask matrix are decomposed into low-rank matrices using matrix factorization, and the decomposition results are weighted and fused to obtain a weighted fused mask matrix.

5. The method for predicting gas concentration in underground coal mine tunneling faces according to claim 4, characterized in that, S433 performs weighted fusion on the decomposition results to obtain a weighted fusion mask matrix. When, this is achieved through formula (2): （2）； In formula (2), and All of these are learnable weight coefficients. Represents the time mask matrix The result of low-rank decomposition. Represents the explicit state dependency matrix The result of low-rank decomposition. T To represent the transpose of a matrix, Indicates the number based on sensor type The generated diagonal matrix, Indicates the number based on historical time points. The generated diagonal matrix, It represents the Kronecker product.

6. The method for predicting gas concentration in underground coal mine tunneling faces according to claim 2, characterized in that, The S45 includes: S451, combine the discretized state vectors corresponding to all types of sensors to obtain the discretized state matrix; S452, combine the historical high-dimensional vectors corresponding to all types of sensors to obtain the historical high-dimensional matrix; S453, based on the state space model, discretized state matrix, weighted fusion mask matrix and historical high-dimensional matrix corresponding to the gas concentration sensor, construct the historical gas concentration feature vector for each historical time node.

7. The method for predicting gas concentration in underground coal mine tunneling faces according to claim 6, characterized in that, S453 is based on the state-space model corresponding to the gas concentration sensor. Discretized state matrix Weighted fusion mask matrix and historical high-dimensional matrix The historical gas concentration feature vector at the t-th historical time node is constructed using formula (3). : （3）； In formula (3), Represents the observation matrix. This indicates the output influence matrix. This represents a vector concatenation operation. The attention pooling operation represents the learnable parameters of the historical gas concentration feature vector at the t-th historical time node, including: and .

8. The method for predicting gas concentration in underground coal mine tunneling faces according to claim 2, characterized in that, S46 uses a multilayer perceptron to extract the historical gas concentration feature vector at the t-th historical time point. Mapped to prediction coefficients When, this is achieved through formula (4): （4）； In formula (4), Describes the multilayer perceptron function. This represents the gated residual activation network function. This represents the learnable weight matrix. This represents a learnable bias vector. This represents the learnable residual weight parameters. The value of n is the second preset quantity.

Images