Dynamic evolution interpolation method for safety monitoring data of fully mechanized coal mining face of coal mine

Abstract

The invention relates to a dynamic evolution interpolation method for safety monitoring data of a fully mechanized coal mining face of a coal mine, and belongs to the field of fully mechanized coal mining face monitoring data interpolation. The method comprises the steps that firstly, an original time sequence is disassembled into a trend part, a seasonal part and a residual error part through a mixed decomposition method, the trend and the seasonal part are extracted through static Loess STL, and the residual error is dynamically adjusted through a Kalman filter; thirdly, modeling trend and seasonal components by using a Gaussian process GPs, and introducing a Shenchang differential equation NODEs to simulate a change rule of a potential state; and finally, constructing a stochastic differential equation (SDE) model, integrating results of the NODEs and the GPs, adding a diffusion term to reflect a random change, solving the SDE to obtain a potential state, and mapping the potential state back to an observation space to generate a predicted value. According to the method, different characteristics in the data are comprehensively considered, and the data with changing time dynamics can be better captured and processed.

Description

Technical Field

[0001] The invention belongs to the field of fully mechanized mining face monitoring data interpolation, and relates to a dynamic evolution interpolation method for safety monitoring data of fully mechanized mining working faces in coal mines. Background Art

[0002] Safety monitoring of coal mine fully-mechanized mining faces is an important part of ensuring coal mine production safety. With the development of automation technology and intelligent sensing technology, a large number of sensors are installed on fully-mechanized mining faces in coal mines to monitor various conditions of the working face in real time, including important parameters such as gas concentration, temperature and humidity, roof pressure, and support status. However, in actual applications, due to factors such as sensor failure and communication interruption, missing values ​​often appear in monitoring data. These missing data will seriously affect the accuracy and reliability of subsequent data analysis and decision-making.

[0003] In order to solve the problem of missing values ​​in the safety monitoring data of coal mine fully mechanized working faces, most of the existing data interpolation methods rely on simple statistical methods or machine learning algorithms, such as mean interpolation, linear interpolation, polynomial interpolation and K nearest neighbor interpolation. Although the traditional method is simple and easy to use, it often does not work well when dealing with time series data with complex dynamic characteristics and nonlinear relationships. It fails to fully consider the potential causal relationship and inherent uncertainty between data points. It cannot effectively deal with the common irregular patterns in coal mine production or sudden changes in time series data. It is not suitable for the heterogeneity of mining data caused by differences in manual operations of different shifts, changes in time periods and the diversity of coal seam structures.

[0004] There are often missing value problems in the safety monitoring data of coal mine fully mechanized working faces. Most of the existing data interpolation methods rely on simple statistical methods or machine learning algorithms, such as mean interpolation, linear interpolation, polynomial interpolation, and K nearest neighbor interpolation. Usually, linear interpolation is the most effective and easy to implement, but it cannot deal with data that is missing from the beginning. Moreover, these methods are not effective when facing time series data with complex dynamic characteristics and nonlinear relationships. They cannot consider the potential connections and inherent uncertainties between data points, and cannot effectively deal with irregular sampling, irregular patterns, or sudden changes in time series data that are common in coal mine production. It is difficult to accurately reflect the true distribution characteristics of data in actual situations, thereby limiting its applicability and accuracy in complex coal mining data analysis. Therefore, it is urgent to develop an interpolation method that can adapt to fully mechanized working face environmental monitoring data. Summary of the invention

[0005] In view of this, the object of the present invention is to provide a dynamic evolution interpolation method for safety monitoring data of fully mechanized mining working faces in coal mines.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A dynamic evolution interpolation method for safety monitoring data of a fully mechanized coal mining face, the method comprising the following steps:

[0008] Perform mixed decomposition on time series data, breaking it down into three parts: trend, seasonality, and residual;

[0009] Use Gaussian processes (GPs) to model trend and seasonal components;

[0010] Model the residual components using Neural Ordinary Differential Equations NODEs;

[0011] Construct a stochastic differential equation (SDE) model, integrate the results of NODEs and GPs, and add diffusion terms to reflect random changes;

[0012] The latent state is obtained by solving the SDE and then mapped back to the observation space to generate the predicted value.

[0013] Furthermore, the hybrid decomposition method includes two steps: static decomposition and dynamic component adjustment.

[0014] Furthermore, the Gaussian process GPs uses a composite covariance function that combines smoothness RBF and periodicity.

[0015] Furthermore, the neural ordinary differential equation NODEs model is based on the idea of ​​encoder-decoder and uses a neural network to parameterize the function of the potential state change rate.

[0016] Furthermore, the stochastic differential equation (SDE) model includes a deterministic part in an ordinary differential equation, a long-term trend term after Gaussian process, a seasonal component after Gaussian process, a short-term correction term updated by a Kalman filter, and a motion diffusion term.

[0017] Furthermore, the solution of SDE is performed step by step using a numerical method.

[0018] The beneficial effects of the present invention are:

[0019] (1) The comprehensive use of models such as neural differential equations, Gaussian processes and stochastic differential equations can more accurately capture and process the complex dynamic characteristics and nonlinear relationships in the safety monitoring data of coal mine fully mechanized mining working faces, thereby improving the accuracy and reliability of interpolation.

[0020] (2) The hybrid decomposition method can effectively process the trend, seasonality and residual components in the data, adapt to the changes in the actual operating conditions of the coal mine working face, and ensure the accuracy of the interpolation results even in the event of sensor failure or communication interruption.

[0021] (3) This method can process different types of safety monitoring data, including important parameters such as gas concentration, temperature and humidity, roof pressure, and support status, and can be flexibly adjusted according to actual needs.

[0022] (4) The present invention provides a complete set of algorithm flows and uses existing open source tools and libraries for implementation, making it convenient for users to quickly deploy and apply them.

[0023] Other advantages, objectives and features of the invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0025] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0026] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0027] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0028] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which 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 operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0029] like Figure 1 As shown, the present invention first uses a hybrid decomposition method to decompose the original time series X(t) into three parts: trend, seasonality and residual, where the trend and seasonality are extracted by static Loess STL, and the residual is dynamically adjusted using the Kalman filter. Then, Gaussian processes (GPs) are used to model the trend and seasonal components, and suitable composite kernel functions (such as RBF and periodic kernels) are selected to capture long-term patterns and their uncertainties. In addition, the Neural Ordinary Differential Equations (NODEs) are introduced to simulate the changing laws of the potential state and provide a smooth and deterministic evolution path. Finally, a stochastic differential equation (SDE) model is constructed to integrate the results of NODEs and GPs, and a diffusion term is added to reflect random changes. The potential state z(t) is obtained by solving the SDE, and is mapped back to the observation space to generate the predicted value X(t). This method comprehensively considers the different characteristics in the data and aims to improve the accuracy of prediction.

[0030] Technical implementation steps:

[0031] (1) Hybrid decomposition of time series data

[0032] The present invention assumes that the time series observation data X(t) can be decomposed, which helps capture stable long-term structure (trend and seasonality) and short-term dynamics (fluctuation and anomaly) by simplifying the complex time series into several manageable components. t , Seasonal ItemS t and the residual term R t The formula is as follows:

[0033] X(t)=T t +S t +R t

[0034] Where X(t) is the observed time series data; T t It is the trend component, representing the long-term trend of data changes; S tis the seasonal component, reflecting the regularity or periodic changes in the data; R t is the residual component, which contains the remainder after trend and seasonality, representing short-term fluctuations and noise.

[0035] The hybrid decomposition method consists of two main steps: static decomposition and dynamic component adjustment. This method balances the static and dynamic methods, making it flexible to handle long-term trends and short-term fluctuations.

[0036] Estimated trend T t , a moving average filter can be applied, and the smoothing formula is as follows:

[0037]

[0038] Where m is the window parameter that controls the smoothness of the trend. The larger m is, the smoother the trend is.

[0039] Seasonal t Usually extracted based on the period average value,

[0040]

[0041] Where P is the period (e.g., 24 hours for daily data) and N is the number of observed periods.

[0042] When T is determined t and S t , the residual component R t The calculation is as follows:

[0043] R t =X(t)-T t -S t

[0044] A dynamic filter is introduced to adjust the residual component R by focusing on short-term fluctuations and anomalies. t The Kalman filter is a recursive estimator that updates its predictions based on new observations. For each time instant t, R t Make predictions This estimate is then corrected as follows:

[0045] Prediction Step: Predict R using the state-space model t The next state. Assume R t Following an autoregressive process:

[0046]

[0047] φ is the coefficient that controls the influence of residual error, w t Yes Obedience Distributed noise.

[0048] Update step: Given a new observation R t , update the prediction based on the observed error:

[0049]

[0050] Where K t is the Kalman gain, which is calculated by the following formula:

[0051]

[0052] here, is the forecast error variance, is the measurement noise variance.

[0053] Updating the residual component through the Kalman filter can better reflect short-term dynamics and anomalies. This approach enables the model to respond to sudden changes or state transitions.

[0054] (2) Modeling the decomposed components (trend and seasonality) of the time series using Gaussian processes

[0055] For the decomposition term f(t) (i.e. the trend term T decomposed in step 1) t , Seasonal ItemS t ) is modeled using a Gaussian process defined by a mean function m(t) and a covariance function k(t,t′).

[0056] f(t)~GP(m(t),k(t,t′))

[0057] Where m(t) = E[f(t)] is the mean function, which represents the expected value of the function at each time t. k(t,t′) = E[(f(t)-m(t))ft′-m(t′) is the covariance function, which encodes the similarity between the function values ​​at different times t and t′. Usually, it is assumed that the data is centered and m(t) = 0. Therefore, GP can be completely defined by the covariance function.

[0058] There are many choices of covariance functions. The present invention selects a composite covariance function that combines smoothness RBF and periodic periodicity, which is defined as follows:

[0059] k(t,t′)=k RBF (t,t′)+k Periodic (t,t′)

[0060]

[0061] in, is the amplitude variance, which is used to control the range of variation of the output value of the Gaussian process regression model. is the length scale and P represents the period.

[0062] Trend T t For example, use the Gaussian process to predict at the new time point t * Trends at:

[0063]

[0064] where y is the vector of observations, T * is the new time point t * The prediction vector of k(t,t) is the covariance matrix between the observation points, k(t * ,t) and k(t,t * ) is the covariance matrix between the observation point and the prediction point, k(t * ,t * ) is the covariance matrix between the prediction points and the prediction points, I represents the identity matrix, is the noise variance of the observations, using To represent noise.

[0065] Given the observed data y, T * The conditional distribution of is as follows:

[0066] T * ∣y~N(μ * ,Σ * )

[0067] where μ * Indicates the predicted trend value at the new moment,

[0068] Σ * represents the uncertainty of trend forecast,

[0069] (3) Modeling continuous-time dynamics in time series using Neural Ordinary Differential Equations

[0070] This step is based on the idea of ​​encoder-decoder. The Neural Ordinary Differential Equation (NODE) ​​framework is a powerful tool for modeling continuous-time dynamics in time series. Its basic principle is to model the latent state z(t) of the time series X(t) as the solution of an ordinary differential equation:

[0071]

[0072] where z(t) is the latent state at time t, f is a neural network parameterized by θ, and θ represents the network parameters learned from the data.

[0073] The NODE model learns a function f that describes the rate of change of the underlying state, enabling it to capture complex temporal patterns. The solution to this ODE at any time t can be obtained by integration:

[0074]

[0075] where z(t0) is the initial state at time t0.

[0076] The latent state z(t) needs to be mapped back to the observed time series X(t). This is usually done using a decoder function h, usually modeled as another neural network:

[0077] X(t)=h(z(t);φ)

[0078] where φ is the parameter of the decoder network.

[0079] To train the model, we need to minimize the predicted time series The difference between the observed value X(t) and the actual observed value X(t). The loss function is usually defined as:

[0080]

[0081] where / / · / / is a suitable norm (e.g., the L2 norm).

[0082] (4) SDE conversion

[0083] The output of the neural differential equation and the Gaussian process is used as the input of the SDE model using the stochastic differential equation (SDE) transformation component. SDE combines the contributions of ODE and GP as well as the noise component. The evolution of the potential state z(t) in SDE is given by the following formula:

[0084] dz(t)=f(z(t),t;θ)dt+T(t)dt+S(t)dt+R(t)dt+g(z(t),t)dW(t)

[0085] Among them, f(z(t), t; θ)dt is the deterministic part of the ordinary differential equation, which describes the smooth evolution of the potential state; T(t)dt is the long-term trend term after Gaussian process; S(t)dt is the seasonal component after Gaussian process; R(t)dt is the short-term correction term updated by the Kalman filter; g(z(t), t)dW(t) is the motion diffusion term, which reflects the accumulated random volatility.

[0086] Finally, a numerical method is used to gradually solve the SDE to obtain the value of z(t) at each time step. For each new observation, the Kalman filter updates R(t) according to the observation data, thereby dynamically adjusting z(t) and then converting the z(t) of each time step into the final predicted value X(t) through h(z(t)).

[0087] (5) Model summary

[0088] The original time series X(t) is decomposed into trend, seasonal and residual components using a hybrid decomposition method. The trend and seasonal components are extracted using static Loess STL, while the residual component is dynamically updated using a Kalman filter to capture short-term fluctuations and anomalies.

[0089] For the trend and seasonal components obtained from the decomposition, Gaussian processes (GPs) with appropriate kernels (e.g., RBF for trend and periodicity for seasonality) are used for modeling to capture the long-term structure and cyclical patterns including uncertainties.

[0090] We exploit Neural Ordinary Differential Equations (NODEs) to model the continuous-time dynamics of the underlying states, providing a smooth, deterministic trajectory that captures complex temporal dependencies and forms deterministic drifts of the SDE.

[0091] A stochastic differential equation (SDE) transformation component is constructed, and the outputs of NODEs and GPs are used as the input of the SDE model. SDE includes a diffusion term to explain random fluctuations. Solving the SDE yields the latent state Z(t), which is then mapped to the observation space to generate the final predicted value X(t).

[0092] The present invention uses the environmental monitoring data of the 1101 fully mechanized mining face of a mine in a certain mining area from September 11, 2024 to September 19, 2024, totaling 13,141 rows and 41 monitoring sensor data (including data of sensors such as oxygen, temperature, humidity, methane, and carbon monoxide). The mask mechanism is used to randomly mask some values ​​in the data set to simulate missing values, and the missing value rate missing_ratio is set to control the missing rate of the data set, and the random seed random_seed is set to ensure that the experiment can be reproduced.

[0093] To verify the effect of the present invention, the random seed random_seed=24 is set, and mean interpolation, linear interpolation, polynomial interpolation and K nearest neighbor interpolation are used for comparative experiments. Mean square error MSE and mean absolute error MAE are used as evaluation indicators, and the smaller the two indicators, the better.

[0094] The calculation formula of MSE is as follows:

[0095]

[0096] Where m is the number of samples, is the difference between the true value and the predicted value.

[0097] The calculation formula of MAE is as follows:

[0098]

[0099] Where m is the number of samples, is the absolute value of the difference between the true value and the predicted value.

[0100] When the missing value rate missing_ratio is set to 20%, on the selected sample data, the MAE and MSE values ​​of mean interpolation, linear interpolation, polynomial interpolation, K nearest neighbor interpolation and the dynamic evolution interpolation algorithm proposed in the present invention are shown in Table 1.

[0101] Table 1

[0102] method MAE MSE Mean interpolation algorithm 202.638 0.679 Linear interpolation algorithm 2.885 0.023 Polynomial interpolation algorithm 3.007 0.029 K nearest neighbor interpolation algorithm 2.898 0.024 Dynamic Evolutionary Interpolation Algorithm 2.901 0.024

[0103] When the missing value rate missing_ratio is set to 30%, on the selected sample data, the MAE and MSE values ​​of mean interpolation, linear interpolation, polynomial interpolation, K nearest neighbor interpolation and the dynamic evolution interpolation algorithm proposed in the present invention are shown in Table 2.

[0104] Table 2

[0105] method MAE MSE Mean interpolation algorithm 304.181 1.033 Linear interpolation algorithm 3.164 0.041 Polynomial interpolation algorithm 3.279 0.049 K nearest neighbor interpolation algorithm 7.368 0.071 Dynamic Evolutionary Interpolation Algorithm 2.947 0.027

[0106] When the missing value rate missing_ratio is set to 40%, on the selected sample data, the MAE and MSE values ​​of mean interpolation, linear interpolation, polynomial interpolation, K nearest neighbor interpolation and the dynamic evolution interpolation algorithm proposed in the present invention are shown in Table 3.

[0107] Table 3

[0108] method MAE MSE Mean interpolation algorithm 383.810 1.329 Linear interpolation algorithm 3.511 0.064 Polynomial interpolation algorithm 3.433 0.057 K nearest neighbor interpolation algorithm 15.693 0.135 Dynamic Evolutionary Interpolation Algorithm 3.109 0.040

[0109] In summary, the method proposed in the present invention has a better effect on data interpolation when the missing value rate of the data set is higher than other methods. When the missing value rate missing_ratio = 20%, the MAE and MSE are 2.901 and 0.024 respectively. When the missing value rate missing_ratio = 30%, the MAE and MSE are 2.947 and 0.027 respectively. When the missing value rate missing_ratio = 40%, the MAE and MSE are 3.109 and 0.040 respectively.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A dynamic evolution interpolation method for safety monitoring data of fully mechanized coal mining working face, characterized by: The method comprises the following steps: Perform mixed decomposition on time series data, breaking it down into three parts: trend, seasonality, and residual; Use Gaussian processes (GPs) to model trend and seasonal components; Model the residual components using Neural Ordinary Differential Equations NODEs; Construct a stochastic differential equation (SDE) model, integrate the results of NODEs and GPs, and add diffusion terms to reflect random changes; The latent state is obtained by solving the SDE and then mapped back to the observation space to generate the predicted value.

2. The method for dynamic evolution interpolation of safety monitoring data of fully mechanized coal mining working face according to claim 1 is characterized in that: The hybrid decomposition includes two steps: static decomposition and dynamic component adjustment.

3. The method for dynamic evolution interpolation of safety monitoring data of fully mechanized coal mining working face according to claim 1 is characterized in that: The Gaussian process GPs uses a composite covariance function that combines a smoothness RBF with a periodic periodic function.

4. The method for dynamic evolution interpolation of safety monitoring data of fully mechanized coal mining working face according to claim 1 is characterized in that: The neural ordinary differential equation NODEs model is based on the idea of ​​encoder-decoder and uses a neural network to parameterize the function of the potential state change rate.

5. The method for dynamic evolution interpolation of safety monitoring data of fully mechanized coal mining working face according to claim 1 is characterized in that: The stochastic differential equation (SDE) model includes a deterministic part in an ordinary differential equation, a long-term trend term after Gaussian process, a seasonal component after Gaussian process, a short-term correction term updated by a Kalman filter, and a motion diffusion term.

6. The method for dynamic evolution interpolation of safety monitoring data of fully mechanized coal mining working face according to claim 1 is characterized in that: The solution to the SDE is performed step by step using a numerical method.

Images