Multi-dimensional prediction method and system for periodic weighting of coal mine working face for edge calculation

Abstract

The invention discloses a multi-dimensional prediction method and system for periodic weighting of a coal mine working face for edge calculation, which are applied to the field of coal mine down pressure early warning and aim at solving the problems that the false alarm rate of the existing weighting early warning is high, the early warning time delay is relatively high, and sufficient time cannot be provided for speed reduction of a coal mining machine, secondary pressure supplement of a hydraulic support, personnel evacuation and the like. According to the method, whether impact noise exists or not is judged by calculating the standard deviation of collected data, when the impact noise exists, online HP filtering processing is carried out on the collected data, LSTM reasoning is carried out on an intermediate frequency fluctuation item after filtering processing, and early warning is carried out based on a reasoning result and a long-term trend item after filtering processing.

Description

A Multidimensional Prediction Method and System for Periodic Pressure of Coal Mine Working Faces Based on Edge Computing Technical Field

[0001] This invention belongs to the field of coal mine downpressure early warning, and specifically relates to a coal mine working face downpressure prediction technology. Background Technology

[0002] Currently, underground coal mines commonly use pressure sensors in the lower cavity of hydraulic support columns to collect working resistance in real time and employ a "fixed threshold method" for periodic pressure alarms: when the pressure exceeds 38 MPa or the safety valve opening rate is >30%, it is considered an incoming pressure. This method is simple and easy to implement, but it is essentially a reactive alarm with an advance warning of less than 10 minutes, failing to provide sufficient time for coal mining machine deceleration, secondary pressure replenishment of hydraulic supports, and personnel evacuation. Some mines have introduced micro-vibration, acoustic emission, or fiber optic grating monitoring to identify incoming pressure through energy accumulation; however, this requires additional drilling to install sensors, resulting in high costs and difficult maintenance. Furthermore, the "energy-incoming pressure" correlation is greatly affected by geological anomalies such as tectonic coal formations and faults, leading to a false alarm rate still exceeding 15%.

[0003] In recent years, some studies have directly input the original support pressure sequence into LSTM for prediction; however, mine pressure signals have dual interference characteristics:

[0004] (1) Long-term drift: caused by the advancement of the working face and the rise in hydraulic oil temperature, which is characterized by a slow upward trend;

[0005] (2) High-frequency noise: originates from pump station pulsations and mechanical vibrations, manifesting as random spikes. Inputting this noise into the LSTM network without preprocessing will result in:

[0006] The network treats drift as a trend, resulting in an overall deviation in predicted values ​​and a peak error greater than 2 MPa.

[0007] High-frequency noise is amplified by LSTM, resulting in false peaks and frequent false alarms;

[0008] It requires deeper networks or longer sequences to converge, has high inference latency, and is difficult to run in real time at the downhole edge.

[0009] Another study combines HP filtering with LSTM for radar echo analysis, but its goal is to separate the uniform motion trend of the target from the micro-Doppler modulation. The HP parameter λ is fixed, there are no physical constraints, and the output is only a classification label. This is fundamentally different from the present invention, which separates the periodic fluctuation of rock fracture, adapts to λ, and outputs a three-dimensional early warning.

[0010] The existing technology has the following problems:

[0011] (1) How to accurately extract the "mid-frequency fluctuation" characteristics that only reflect the periodic fracture of rock strata from the original support pressure containing drift and noise?

[0012] (2) How to achieve an inference latency of ≤30 ms on ARM edge chips such as RK3568 to meet the online prediction requirements of a 5 s sampling period?

[0013] (3) How to increase the lead time of the warning from <10 minutes to ≥30 minutes, while ensuring that the peak prediction RMSE is reduced by more than 40%?

[0014] Therefore, there is an urgent need for a technical solution that requires no additional hardware, can run in real time at the edge, and can accurately predict the cycle more than 30 minutes in advance to reduce time and intensity. Summary of the Invention

[0015] To address the aforementioned technical problems, this invention proposes a multidimensional prediction method and system for the periodic pressure of coal mine working faces based on edge computing.

[0016] One of the technical solutions adopted in this invention is: a multi-dimensional prediction method for periodic pressure on coal mine working faces oriented towards edge computing, comprising:

[0017] S1. Collect electro-hydraulic control pressure data of each support based on pressure sensors;

[0018] S2. Calculate the mean and standard deviation of the current pressure data. If the standard deviation is greater than the set threshold, the current pressure data is considered to have impact noise.

[0019] S3. Perform online HP filtering on the current pressure data containing impact noise to obtain the long-term trend term and the mid-frequency fluctuation term;

[0020] S4. Input the intermediate frequency fluctuation term obtained by filtering into the LSTM network to predict the intermediate frequency fluctuation in the future.

[0021] S5. Calculate the long-term trend term for future time based on the long-term trend term obtained in step S3;

[0022] S6. Based on the mid-frequency fluctuations of future time predicted by the LSTM network and the long-term trend term of future time obtained in step S5, the reconstruction curve is obtained.

[0023] S7. Identify the global peak of the reconstructed curve and its corresponding lead time, and issue an early warning.

[0024] The second technical solution adopted in this invention is: a multi-dimensional prediction system for periodic pressure of coal mine working faces oriented to edge computing, comprising: a first data acquisition module, a second data acquisition module, an online HP filtering module, an LSTM network prediction module, a third data acquisition module, a reconstructed curve construction module, a fourth data acquisition module, and an early warning module;

[0025] The first data acquisition module is used to obtain the mean and standard deviation of the current data;

[0026] The second data acquisition module is used to acquire data whose standard deviation is greater than a threshold.

[0027] The online HP filtering module is used to filter the data acquired by the second data acquisition module to obtain the long-term trend term and the mid-frequency fluctuation term;

[0028] The LSTM network prediction module takes the intermediate frequency fluctuation term as input and predicts the intermediate frequency fluctuation term for future times.

[0029] The third data acquisition module obtains the long-term trend term for future time based on the long-term trend term obtained by the online HP filtering module;

[0030] The reconstruction curve construction module constructs the reconstruction curve based on the mid-frequency fluctuation term of the future time output by the LSTM network prediction module and the long-term trend term of the future time obtained by the third data acquisition module.

[0031] The fourth data acquisition module obtains the global peak value and lead time of the reconstructed curve;

[0032] The early warning module issues warnings based on the global peak value and lead time obtained by the fourth data acquisition module.

[0033] The beneficial effects of this invention: This invention provides an online prediction method and system for coal mine periodic pressure using HP filtering and LSTM fusion. It reuses existing electro-hydraulic control pressure data from support structures and achieves the following effects with zero cost and zero downtime through a technology chain of "adaptive HP detrending + lightweight LSTM fluctuation prediction + physically constrained three-dimensional criteria":

[0034] It provides early warning of pressure 30–120 minutes in advance; outputs three-dimensional early warning information in terms of time, intensity, and confidence; edge inference latency is ≤30 ms; and it is self-updating throughout its entire lifecycle without the need for manual parameter tuning.

[0035] The present invention has the following advantages:

[0036] (1) The first "adaptive HP + lightweight LSTM" serial architecture is designed specifically for the triple characteristics of "drift + noise + mutation" in mining pressure, rather than a simple porting of HP+LSTM in other fields such as radar / finance;

[0037] (2) λ is quantitatively calculated according to the sampling rate, realizing "one-click adaptation" for different working surfaces without the need for expert parameter adjustment;

[0038] (3) Three-dimensional physical criteria significantly reduce false alarms and achieve zero pressure rack accidents;

[0039] (4) Fully reuse existing stress data, with zero new hardware, zero installation downtime, and deployment costs approaching zero;

[0040] (5) Edge inference ≤ 30 ms, which meets the downhole real-time requirements, while the latency of the general LSTM solution is > 200 ms;

[0041] (6) The self-updating mechanism ensures that the model remains effective as geological conditions evolve, solving the industry problem of "model degradation". Attached Figure Description

[0042] Figure 1 is a flowchart of a multi-dimensional prediction method for the periodic pressure of a coal mine working face based on edge computing according to the present invention.

[0043] Figure 2 is a schematic diagram of a multi-dimensional prediction system for the periodic pressure of a coal mine working face based on edge computing according to the present invention. Detailed Implementation

[0044] As shown in Figure 1, the multi-dimensional prediction method for periodic pressure in coal mine working faces based on edge computing according to the present invention includes the following steps:

[0045] S1. Data Acquisition: This invention provides a parallel T-junction at the original M14×1.5 pressure ball valve in the lower cavity of each support column, connecting a pressure sensor for data acquisition. The connected pressure sensor does not require pressure relief. The pressure sensor is powered by a 24V electro-hydraulic intrinsically safe power supply (maximum output 1A) from the support, with a total current <60mA. MHYVP 1×4×7 / 0.52 mining shielded cable is used to connect the 128 pressure sensors installed in this embodiment to the edge gateway via two RS485 buses in a daisy-chain configuration. The baud rate is 256 kbps, the refresh cycle is 200 ms, and the data is packetized into 5-second averages and transmitted within the gateway.

[0046] In this embodiment, the pressure sensor used is the intrinsically safe mining pressure sensor GPD60W, with a pressure rating of 0–60 MPa, an accuracy of ≤0.5%, and a resolution of 0.1 MPa.

[0047] In this embodiment, the edge gateway is an underground explosion-proof and intrinsically safe edge computing gateway Exd[ib]IMb, with a main controller RK3568, 4 X A55 1.8 GHz, and 2 GB LPDDR4.

[0048] This embodiment uses an underground gigabit ring network with TCP / IP + MQTT 3.1.1, QoS=1, and a 48-hour disconnection buffer; the ground server is an Intel Xeon Silver 4310, 128 GB RAM, and 2×RTX 4080, responsible for model retraining and alarm push.

[0049] S2. Use libmodbus for asynchronous scanning with a circular buffer of 30 s. Calculate the mean μ and standard deviation σ every 5 s. If σ > 0.5 MPa, it is considered impulse noise and abnormal samples are removed using Hampel filtering.

[0050] S3. Online HP filtering: Maintain the original sliding sequence (2880 points) in memory for 4 hours, and solve the tridiagonal equation using the Thomas algorithm with a complexity of O(n).

[0051] This invention uses adaptive HP filtering to accurately extract the period to suppress fluctuation characteristics; unlike the existing HP+LSTM, this invention dynamically adapts λ to the coal mining advance rhythm, and its physical meaning is "rock stratum fracture period scale".

[0052] 4-hour sliding window pressure sequence Perform HP filtering:

[0053] ;

[0054] in, Represents the original observation sequence The vector formed , , This represents the number of original observation sequences within a 4-hour sliding window, such as 2880 points. Sampling occurs once every 5 seconds, meaning that a 4-hour sliding window will collect 2880 original observation sequences. Indicates the long-term trend term. Mid-frequency fluctuation term; This represents the smoothing parameter, which controls the trend term. The smoothness; is a (n−2)×n second-order difference matrix, which is used to perform discrete second-order difference operations on the trend term g to measure the local curvature of g; I is an n × n identity matrix, that is, an n-order square matrix with 1s on the main diagonal and 0s on the rest. This represents the transpose operator.

[0055] Introducing a multi-physical parameter coupling λ calculation formula, and incorporating propulsion velocity. and coal seam thickness As a key influencing factor, it is tied to the specific working conditions of the coal mine face, rather than general signal processing.

[0056] The smoothing parameter λ is given by the following multi-parameter coupling formula:

[0057]

[0058] Among them, f s This indicates the sampling frequency of the pressure data, such as 0.2 Hz, which means once every 5 seconds; The reference sampling frequency is 1 Hz. This indicates the daily advance speed of the working face, which is obtained in real time by the displacement sensor of the coal mining machine. This indicates the current coal seam thickness at the working face, dynamically mapped from the geological exploration model; m is the unit "meter" and d is the unit "day"; This represents an empirical coefficient, with an initial value of [value missing]. This involves using historical data to inversely calibrate events. The formula uses f... s , , Perform coupling with unit normalization separately.

[0059] Periodicity of rock strata fracture Proportional, that is The signal period range preserved by the HP filter should match... The physical meaning is clear. This formula is only applicable to the coal mine cycle pressure scenario coupled with the "advance speed-coal thickness-sampling rate" of this invention;

[0060] At the edge layer, the original stent pressure sequence over the past 4 hours (2880 points) was analyzed. Performing HP filtering yields:

[0061]

[0062] in This represents the long-term trend (drift). This represents the mid-frequency fluctuation term (rock fracture signal). Solving this optimization problem is equivalent to solving a tridiagonal linear system. The Thomas algorithm is used on an RK3568 platform and can be completed in real-time with O(n) complexity. The solution is obtained... According to the formula Calculated Output It serves as the sole input to the LSTM.

[0063] Ensure one-click adaptation of working surfaces with different sampling rates (5s / 1min), retaining 0.5–8.

[0064] The different sampling rates (5s / 1min) for the working face here specifically refer to the fact that some working faces may use 1 minute of sampling, while others may use 5s of sampling. In this embodiment, 5s is used.

[0065] The one-click adaptation here should be understood as: no manual intervention or parameter readjustment is required. The system can automatically and dynamically adjust the smoothing parameter λ based on parameters such as the sampling rate, advance speed, and coal seam thickness of the current working face, thereby ensuring that the HP filter can always effectively separate the "mid-frequency fluctuation" characteristics that reflect rock strata fractures.

[0066] The "0.5–8" here should be understood as: the "intermediate frequency fluctuation term" retained by the HP filter. The corresponding cycle range is measured in minutes (min). Based on field experience in coal mines, a complete "cycle pressure" process typically lasts between 0.5 and 8 minutes. This range is derived from statistics of a large number of historical pressure events and covers the vast majority of real-world scenarios.

[0067] S4, LSTM inference:

[0068] Network structure: input(1440,1)→LSTM(128, return_seq=False)→Dropout(0.2)→Dense(720)→output(720,1);

[0069] The weight file is 2.1 MB, with a loading time of 180 ms and a single inference time of 28 ms (RK3568 single thread).

[0070] Input normalization:

[0071]

[0072] Updated on a rolling basis using data from the past 24 hours;

[0073] The output, after inverse normalization, yields ĉ.

[0074] In this embodiment, the pure wave sequence output from step S2 is filtered. As the sole input to the LSTM, 1440 points are used as the input in this embodiment. Output the future 720 points .

[0075] S5, Trend Extrapolation; using a "local linear + damped" model:

[0076]

[0077] Among them, damping factor To prevent long-term drift, k represents the sequence number of the future time step (or the number of prediction steps), which is an integer variable that starts from 1 and increments, representing the prediction of the kth time point in the future from the current time n; for example, if the sampling period Δt = 5 seconds, then k = 1 predicts 5 seconds later; k = 2 predicts 10 seconds later; k = 720 predicts 60 minutes later.

[0078] This embodiment shows the trend item. This is mainly caused by the continuous loading of overburden due to the advancement of the working face, and its slope is related to the advancement speed. and coal seam hardness Positive correlation. Therefore, the trend extrapolation is defined as:

[0079]

[0080] in: The daily advance rate of the working face under the baseline working conditions. The coal seam hardness under the baseline operating conditions; The baseline drift rate is determined through statistical analysis of historical data. This is the damping coefficient, used to prevent over-prediction in the long term; Real-time data is obtained from the position sensors of the coal mining machine. Loaded from the geological exploration report.

[0081] S6, Reconstruction Curve:

[0082] ;

[0083] The long-term prediction drift problem was completely solved, and the RMSE was reduced from 1.8 MPa to 0.86 MPa.

[0084] Find the global peak value P max =max( ) and corresponding lead time t max =k max ×Δt;K max This represents the time step index corresponding to the global peak on the predicted reconstruction curve, and its value is obtained by traversing all values ​​in the range k=1 to 720. The value is obtained by taking the position of the maximum value; Δt is the data sampling period of the system, which is 5 seconds in this embodiment. This value is determined by the data packaging strategy of the edge gateway.

[0085] S7. Early Warning Decision-Making:

[0086] Unlike other LSTM-based mine pressure studies, which rely on a single threshold (e.g., >38 MPa), this invention introduces a combined geological and operational condition constraint, reducing the false alarm rate from 15% to 0%. An alert is triggered only when the following three conditions are met simultaneously:

[0087] Dynamic intensity threshold:

[0088]

[0089] Suitable for mining heights of 3.5–8.8 m, with automatic threshold increase for ultra-high mining heights.

[0090] Time window:

[0091]

[0092] Dynamic step range:

[0093]

[0094] The faster the advance, the larger the step distance.

[0095] When all criteria are met, an alert is triggered:

[0096] ① (Strength criterion);

[0097] ② (Time criterion);

[0098] ③ The distance ΔL from the last press advance ∈ .

[0099] in, This indicates the estimated number of minutes until the next periodic pressure event occurs. ΔL represents the distance the working face has advanced since the last periodic pressure event, a key operating parameter well-known in the field.

[0100] Figure 2 shows a schematic diagram of the system of the present invention. The present invention is a multi-dimensional prediction system for periodic pressure of coal mine working face oriented to edge computing, comprising: a first data acquisition module, a second data acquisition module, an online HP filtering module, an LSTM network prediction module, a third data acquisition module, a reconstructed curve construction module, a fourth data acquisition module, and an early warning module.

[0101] The first data acquisition module is used to obtain the mean and standard deviation of the current data;

[0102] The second data acquisition module is used to acquire data whose standard deviation is greater than a threshold.

[0103] The online HP filtering module is used to filter the data acquired by the second data acquisition module to obtain the long-term trend term and the mid-frequency fluctuation term; in this embodiment, it is specifically applied to a 4-hour sliding window pressure sequence. Perform HP filtering: λ is an intermediate parameter coupled with multiple physical parameters, dynamically adapting to the coal mining advance rhythm, and its physical meaning is "the periodic scale of rock strata fracture".

[0104] The LSTM network prediction module takes the intermediate frequency fluctuation term as input and predicts the intermediate frequency fluctuation term for future times.

[0105] The third data acquisition module obtains the long-term trend term for future time based on the long-term trend term obtained by the online HP filtering module;

[0106] The reconstruction curve construction module constructs the reconstruction curve based on the mid-frequency fluctuation term of the future time output by the LSTM network prediction module and the long-term trend term of the future time obtained by the third data acquisition module.

[0107] The fourth data acquisition module obtains the global peak value and lead time of the reconstructed curve;

[0108] The early warning module issues warnings based on the global peak value and lead time obtained by the fourth data acquisition module.

[0109] This invention employs an edge-cloud collaborative self-update mechanism:

[0110] False alarm tags are transmitted back from the edge.

[0111] Automatic daily retraining of LSTM in the cloud (Optuna parameter tuning + ONNX export);

[0112] The edge model, which uses MQTT OTA hot updates, achieves zero downtime and zero human intervention.

[0113] The early warning effect of the method of this invention is shown in Table 1. Compared with the early warning time of less than 10 minutes of the traditional method, the method of this invention can achieve an early warning time of more than 30 minutes, providing sufficient response window for key safety operations such as coal mining machine speed reduction, secondary pressure replenishment of supports, and personnel evacuation, significantly improving the safety of underground operations. The low RMSE in Table 1 reflects the high-precision peak prediction capability of this invention. The pressure peak prediction error is extremely small, indicating that the reconstructed curve of this invention closely approximates the actual pressure intensity, effectively avoiding excessive intervention or insufficient protection due to misjudgment of intensity. The zero false alarm rate reflects that this invention effectively suppresses false alarms through "adaptive HP filtering + three-dimensional physical criteria (intensity-time-step distance)," solving the long-standing pain point of high false alarms in the industry and improving the reliability of the system.

[0114] Table 1. Early warning effect of the method of the present invention

[0115]

[0116] Those skilled in the art will recognize that the embodiments described herein are for the purpose of helping to understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A multi-dimensional prediction method for periodic pressure on coal mine working faces oriented towards edge computing, characterized in that, include: S1. Collect electro-hydraulic pressure data of each support based on pressure sensors, and obtain a pressure sequence each time; S2. Calculate the mean and standard deviation of the current pressure sequence. If the standard deviation is greater than the set threshold, it is considered that the current pressure sequence has impact noise; S3. Perform online HP filtering on the current pressure sequence with impact noise to obtain the long-term trend term and the mid-frequency fluctuation term. S4. Input the intermediate frequency fluctuation term obtained by filtering into the LSTM network to predict the intermediate frequency fluctuation in the future. S5. Calculate the long-term trend term for future time based on the long-term trend term obtained in step S3; S6. Based on the mid-frequency fluctuations of the future time predicted by the LSTM network and the long-term trend term of the future time obtained in step S5, the reconstruction curve is obtained; S7. Find the global peak of the reconstruction curve and the corresponding lead time, and issue an early warning.

2. The multidimensional prediction method for periodic pressure of coal mine working faces oriented towards edge computing according to claim 1, characterized in that, The expression for online HP filtering of the current pressure sequence containing impact noise is: ;in, Indicates the current pressure sequence. Indicates the smoothing parameter. For long-term trend items, This is a mid-frequency fluctuation term. Represents a second-order difference matrix. Represents the identity matrix. This indicates transpose.

3. The multidimensional prediction method for periodic pressure of coal mine working faces oriented towards edge computing according to claim 2, characterized in that, The smoothing parameter λ is the result of introducing the coupling of multiple physical parameters, and is expressed as: ; where f s Indicates the sampling frequency of pressure data. Indicates the daily progress rate of the working face. This indicates the current coal seam thickness at the working face. Represents the empirical coefficient. The reference sampling frequency is m, where m is the unit "meter" and d is the unit "day".

4. The multidimensional prediction method for periodic pressure of coal mine working faces based on edge computing as described in claim 3, characterized in that, Periodicity of rock strata fracture Proportional.

5. A multi-dimensional prediction method for periodic pressure of coal mine working faces based on edge computing as described in claim 4, characterized in that, Step S5 specifically uses the "local linear + damped" model to calculate the long-term trend term in future time, and the calculation formula is as follows: ;in, This indicates the daily advance rate of the working face under the baseline operating conditions. This indicates the coal seam hardness under baseline operating conditions. 、 ； As the reference drift rate, The damping coefficient is... This refers to the hardness of the coal seam.

6. A multi-dimensional prediction method for periodic pressure of coal mine working faces based on edge computing, as described in claim 5, is characterized in that... The warning in step S7 includes the following process: calculating the dynamic intensity threshold: ;Calculation time window: ; Calculate the dynamic step range: An alarm will be triggered when the following conditions are met simultaneously: the global peak value is greater than or equal to the dynamic intensity threshold. The distance from the last press advance is within the dynamic step range.

7. A multi-dimensional prediction system for periodic pressure in coal mine working faces oriented towards edge computing, characterized in that, include: The system comprises a first data acquisition module, a second data acquisition module, an online HP filtering module, an LSTM network prediction module, a third data acquisition module, a reconstructed curve construction module, a fourth data acquisition module, and an early warning module; the first data acquisition module is used to acquire the mean and standard deviation of the current data; the second data acquisition module is used to acquire data whose standard deviation is greater than a threshold. The online HP filtering module is used to filter the data acquired by the second data acquisition module to obtain the long-term trend term and the mid-frequency fluctuation term; the LSTM network prediction module uses the mid-frequency fluctuation term as input to predict the mid-frequency fluctuation term for future time. The third data acquisition module obtains the long-term trend term for future time based on the long-term trend term obtained by the online HP filtering module; The reconstruction curve construction module constructs the reconstruction curve based on the mid-frequency fluctuation term of the future time output by the LSTM network prediction module and the long-term trend term of the future time obtained by the third data acquisition module; the fourth data acquisition module obtains the global peak value and lead time of the reconstruction curve. The early warning module issues warnings based on the global peak value and lead time obtained by the fourth data acquisition module.

8. A multi-dimensional prediction system for periodic pressure in coal mine working faces oriented towards edge computing, as described in claim 7, is characterized in that... The online HP filtering module is used to filter the data acquired by the second data acquisition module. The expression is: ;in, Indicates the current pressure sequence. Indicates the smoothing parameter. For long-term trend items, This is a mid-frequency fluctuation term. Represents a second-order difference matrix. Represents the identity matrix. This indicates transpose.

9. A multi-dimensional prediction system for periodic pressure in coal mine working faces oriented towards edge computing, as described in claim 8, is characterized in that... The smoothing parameter λ is the result of introducing the coupling of multiple physical parameters, and is expressed as: ; where f s Indicates the sampling frequency of pressure data. Indicates the daily progress rate of the working face. This indicates the current coal seam thickness at the working face. Represents the empirical coefficient. The reference sampling frequency is m, where m is the unit "meter" and d is the unit "day".

Images