An environment early warning system based on data processing

Abstract

The application discloses an environment early warning system based on data processing and belongs to the technical field of safety monitoring, comprising a sensing layer, a data fusion edge and storage layer, a prediction and early warning layer and an application layer; the sensing layer is used for collecting original data through a plurality of sensors; the data fusion edge and storage layer are used for assembling and processing the original data of the sensors; the prediction and early warning layer is used for predicting a step distance prediction value and a final intensity grade according to the original data after the assembly and processing; and the application layer is used for pushing the step distance prediction value and the final intensity grade. The application can automatically push an alarm and link to measures such as pump station pressure limiting, tunneling speed reduction and sound-light alarm, forms a closed loop of "early warning-> alarm-> control-> feedback" in cooperation with on-site feedback, and significantly improves the mine safety response speed and treatment effect.

Description

Technical Field

[0001] This invention belongs to the field of safety monitoring technology, specifically relating to an environmental early warning system based on data processing. Background Technology

[0002] Traditional pressure monitoring mainly relies on physical sensors to monitor the environment in real time in order to obtain information such as support resistance and roof condition.

[0003] By uploading data to a cloud-based intelligent analysis system using IoT technology, key values ​​such as support resistance, roof delamination, and mining stress can be visualized in real time and automatically generated as early warning reports.

[0004] It should be noted that traditional monitoring methods mostly rely on single-indicator alarms, and wired monitoring is greatly affected by the progress of work, often facing problems such as inconvenient deployment and insufficient coverage.

[0005] Currently, intelligent forecasting and early warning systems still face several challenges in practical applications. First, the volume and quality of monitoring data are enormous: sensors are susceptible to noise and data loss due to vibration and humidity, making data preprocessing difficult; improper handling of missing or abnormal data can reduce prediction accuracy. Second, changes are influenced by geological conditions, mining processes, and other factors, and the data exhibits significant spatiotemporal correlations, making it difficult for a single model to fully capture its complex characteristics. Differences in data format and physical meaning between different monitoring devices also pose challenges to data fusion. Furthermore, model training and deployment must overcome real-time requirements and resource constraints: deep neural networks have numerous parameters, requiring extensive historical data and computational resources for training, while network bandwidth and computing power are limited, creating technical bottlenecks for real-time online prediction. Moreover, current big data and large-scale models are essentially "black boxes," lacking interpretability, and the perception of the reliability and decision-making basis of model predictions is limited, requiring further improvement in engineering applications.

[0006] Finally, due to significant differences in geological environments, the generalization ability of the same prediction model is often insufficient, requiring targeted parameter adjustments or retraining. In summary, data acquisition and quality control, model generalization and interpretability, multi-source data fusion, and system real-time performance are key issues that urgently need to be addressed in current intelligent mine pressure prediction. Summary of the Invention

[0007] To address the problems of insufficient multi-source data fusion, poor real-time performance and accuracy, low level of prediction intelligence, and low system integration in existing monitoring and early warning systems for mine pressures such as fully mechanized mining faces, roadway surrounding rock, roof delamination, and rockbursts, this invention proposes an environmental early warning system based on data processing.

[0008] The technical solution of the present invention is: an environmental early warning system based on data processing, comprising a perception layer, a data fusion edge and storage layer, a prediction and early warning layer, and an application layer;

[0009] The perception layer is used to collect raw data through several sensors;

[0010] The data fusion edge and storage layer are used to assemble and process the raw data from the sensors;

[0011] The prediction and early warning layer is used to predict the step distance and the final intensity level based on the assembled and processed raw data;

[0012] The application layer is used to push the step distance prediction value and the final intensity level.

[0013] Furthermore, the data fusion edge and storage layer assemble and process the raw sensor data, including the following steps:

[0014] A1. Determine the measurement range based on the communication protocols and raw data of each sensor;

[0015] A2. Perform self-tests according to the communication protocols of each sensor, retain the sensors with normal self-test status, and proceed to A3.

[0016] A3. Compare the sensor's raw data with its range, retain the sensors whose raw data is within the range, and proceed to A4.

[0017] A4. If the sensor is mounted on a bracket, assemble the sensor's motion information into JSON format data; otherwise, assemble the sensor's motion information into a dictionary.

[0018] Furthermore, the prediction step size and final intensity level of the early warning layer include the following steps:

[0019] B1. Obtain the time window based on the raw data after assembly. The internal pressure, initial support force, and final resistance, among which, Represents a timestamp. This indicates the length of time it takes to sample backwards from the current moment;

[0020] B2. Normalize the pressure, initial support force, and final resistance within the time window;

[0021] B3. Normalize the top-level off-board data;

[0022] B4. Based on the normalized initial support force and normalized pressure, calculate the correlation between the initial support force and historical pressure fluctuations.

[0023] B5. Based on the normalized final resistance and normalized pressure, calculate the correlation between the final resistance and historical pressure fluctuations.

[0024] B6. Calculate the correlation between the initial support force and the delamination of the top plate based on the normalized initial support force and the normalized delamination.

[0025] B7. Calculate the correlation between the final resistance and the roof delamination based on the normalized final resistance and the normalized delamination.

[0026] B8. Construct feature vectors based on the correlation between initial support force and historical pressure fluctuations, the correlation between final resistance and historical pressure fluctuations, the correlation between initial support force and roof delamination, and the correlation between final resistance and roof delamination.

[0027] B9. Obtain the hidden state based on the feature vector;

[0028] B10. Perform weighted fusion on the hidden states to obtain a weighted feature vector;

[0029] B11. Obtain the step size prediction value based on the weighted feature vector;

[0030] B12. Generate the final intensity level;

[0031] B13. Use the total loss function to train and adjust the step size prediction and the final intensity level.

[0032] Furthermore, in B2, the expression for normalizing the pressure is:

[0033] ;

[0034] ;

[0035]

[0036] in, This represents the normalized pressure. Indicates the first The original value of the support pressure at all times. This represents the arithmetic mean of pressure within a time window. This represents the standard deviation of pressure values ​​within a time window. This represents the width of the sliding window used for normalization. Represents a timestamp. Indicates the first The raw pressure values ​​collected at any time;

[0037] In B2, the expression for normalizing the initial support force is:

[0038] ;

[0039] in, This represents the initial support force after normalization. Record the current pressure sensor value at the moment when the pressure stabilizes and is confirmed.

[0040] In B2, the expression for normalizing the final resistance is:

[0041] ;

[0042] in, This represents the final resistance after normalization. This indicates the pressure value when the support frame is moved.

[0043] In B3, the expression for normalizing the top-level off-board data is:

[0044] ;

[0045] ;

[0046] in, This indicates the separation of layers after normalization. Indicates the average value of the separation layer. Indicates recent Mean depth of point delamination, Indicates a point in time Data on top slab delamination at that time Indicates a relative time index. Indicates recent Standard deviation of the point separation depth.

[0047] Furthermore, in B4, the degree of correlation between initial support force and historical pressure fluctuations... The expression is:

[0048] ;

[0049] in, This represents the initial support force after normalization. This indicates a deviation in historical time series pressure. Indicates the window length. Indicates a relative time index;

[0050] In B5, the correlation between the final resistance and historical pressure fluctuations. The expression is:

[0051] ;

[0052] in, This represents the final resistance after normalization;

[0053] In B6, the correlation between the initial support force and the delamination of the top slab is significant. The expression is:

[0054] ;

[0055] In B7, the correlation between final resistance and top delamination. The expression is:

[0056] ;

[0057] in, This indicates the separation of layers after normalization.

[0058] Furthermore, in B8, the feature vector The expression is:

[0059] ;

[0060] in, Table Normalize stress values ​​at all times. Table Normalize stress values ​​at all times. Indicates in Normalize stress values ​​at all times. Represents a sequence of pressure increments. Represents a binary sequence of stent movements. Indicates real-time progress. Indicates geological structure parameters, This represents the initial support force after normalization. This represents the final resistance after normalization. This indicates the separation of layers after normalization. Indicates in The increment of delamination at the top slab at any given time. Indicates the pressure gradient between adjacent supports. Indicates the degree of coupling between pressure and action. Indicates the degree of coupling between pressure and propulsion. Indicates in The increment of delamination at the top slab at any given time. This indicates the degree of correlation between initial support force and historical pressure fluctuations. This indicates the degree of correlation between the final resistance and historical pressure fluctuations. This indicates the degree of correlation between the initial support force and the delamination of the top slab. This indicates the degree of correlation between the final resistance and the top plate delamination.

[0061] Furthermore, in B9, the hidden state The expression is:

[0062] ;

[0063] in, This represents a gated loop network. Represents the eigenvector. The parameters represent the gated recurrent network. Represents a real matrix. Indicates a time step. This indicates the length of the output vector at each time point;

[0064] In B10, the weighted eigenvectors The expression is:

[0065] ;

[0066] ;

[0067] in, This indicates the transpose operation. Represents the attention weight vector. This represents the activation function. Indicates attention weight;

[0068] In B11, the step size prediction value The expression is:

[0069] ;

[0070] in, Represents the regression weight matrix. Indicates the regression bias term;

[0071] Furthermore, B12 includes the following sub-steps:

[0072] B121. Based on the normalized pressure, the multi-channel time series characteristic sequence is obtained;

[0073] B122. Input the multi-channel time-series feature sequence into a bidirectional LSTM;

[0074] B123. Based on the state vector output by the bidirectional LSTM, obtain the probability value for each intensity level;

[0075] B124. Based on the probability value of each intensity level, the final intensity level is obtained.

[0076] Furthermore, in B121, the multi-channel time-series feature sequence The expression is:

[0077] ;

[0078] in, Represents the normalized pressure sequence. Indicates the kernel size. Indicates the number of convolution kernels. Indicates the step size. This represents a one-dimensional convolutional neural network;

[0079] In B122, the state vector output by the bidirectional LSTM The expression is:

[0080] ;

[0081] in, This represents the hidden unit dimension of a bidirectional LSTM. Indicates a bidirectional LSTM;

[0082] In B123, the probability value for each intensity level. The expression is:

[0083] ;

[0084] in, This represents the activation function. Indicates the first Input data Representation vector, Represents a linear weight matrix. Indicates the bias term;

[0085] In B124, the final intensity level The expression is:

[0086] ;

[0087] in, This represents the set of independent variables that cause the function to reach its maximum value;

[0088] In B13, the total loss function The expression is:

[0089] ;

[0090] ;

[0091] ;

[0092] in, The weights represent the regression task loss. The weights represent the loss of the classification task. This indicates the loss of the return mission. Indicates the loss of the classification task. Indicates the first The step size predicted by the sample model. Indicates the first The actual step distance of each sample. Indicates the first The coding of the intensity label of the sample, The model predicts the first... The sample belongs to the intensity category The probability, Represents the total number of samples. Represents a logarithmic function.

[0093] The beneficial effects of this invention are:

[0094] (1) The present invention performs real-time edge preprocessing and standardization. The perception layer and the data fusion edge layer complete filtering, verification and standardization at the acquisition end, which greatly reduces the data transmission delay and noise interference of mine pressure data, so that the prediction model input always maintains high-quality, low-latency real-time data.

[0095] (2) This invention ensures data consistency and reliability through a multi-layer architecture. It adopts a layered design of perception layer, data fusion and storage layer, prediction and early warning layer and application layer. Each layer has a clear responsibility. Through middleware persistence and policy push, the data link is ensured to be reliable from end to end, avoiding single point of failure.

[0096] (3) This invention adopts an advanced hybrid neural network model. The step size prediction uses GRU and attention mechanism. The intensity judgment combines one-dimensional convolutional neural network, bidirectional LSTM and activation function classification, which can capture local mutations and full sequence dependencies at the same time. It has higher accuracy and robustness than traditional threshold or single model.

[0097] (4) The present invention features online adaptive closed-loop learning. The system updates only the parameters of the attention layer and the output layer through real-time error detection and incremental fine-tuning, ensuring that the model can quickly adapt to changes in the mining pressure environment and effectively reducing the performance degradation caused by long-term drift.

[0098] (5) The application layer of the present invention provides automated linkage control. When the prediction result triggers the early warning condition, the application layer can automatically push an alarm and link measures such as pump station pressure limiting, tunneling deceleration, and audible and visual alarms. Combined with on-site feedback, it forms a closed loop of "early warning → alarm → control → feedback", which significantly improves the mine safety response speed and management effect. Attached Figure Description

[0099] Figure 1 This is a schematic diagram of the structure of an environmental early warning system based on data processing. Detailed Implementation

[0100] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0101] like Figure 1As shown, the present invention provides an environmental early warning system based on data processing, including a perception layer, a data fusion edge and storage layer, a prediction and early warning layer, and an application layer;

[0102] The perception layer is used to collect raw data through several sensors;

[0103] The data fusion edge and storage layer are used to assemble and process the raw data from the sensors;

[0104] The prediction and early warning layer is used to predict the step distance and the final intensity level based on the assembled and processed raw data;

[0105] The application layer is used to push the step distance prediction value and the final intensity level.

[0106] This invention is based on a fusion analysis and prediction model of multi-source time-series data, including support pressure sensors, gas monitoring data, travel sensors, geological information, and roof delamination data, to achieve real-time monitoring and risk warning of dynamic changes in mine pressure.

[0107] To ensure system data source timestamp synchronization, an NTP time synchronization mechanism is uniformly adopted to ensure accurate alignment of timestamps from all sensors, with a time precision of milliseconds. All timestamps use UTC time, with the time format: yyyy-MM-ddTHH:mm:ss.fffZ. For example, 2025-07-02T10:15:32.150Z represents 10:15:32.150 milliseconds on July 2, 2025.

[0108] In the sensing layer, various sensors are scientifically deployed based on the layout of the mine roadways, the fully mechanized mining face, the key support structures and geological conditions, including: support pressure sensors, stroke displacement sensors, roof delamination sensors, and environmental sensors (such as gas, temperature and humidity).

[0109] Each sensor is assigned a unique device ID and installation location identifier to form a complete basic data acquisition. Each sensor is initially calibrated using a dedicated tool to ensure the range, sensitivity, and communication parameters are met. The sensor information of each sensor is then stored in the device configuration database in the following format: [Device ID], [Sensor Type], [Range], [Sensitivity], [Communication Parameters], [Installation Location].

[0110] The sensor continuously collects the raw physical quantities of the sensor at a set sampling frequency. The sensor's built-in hardware module automatically performs local preprocessing, including:

[0111] Initial data screening: Remove erroneous data that clearly exceeds the physical reasonable range;

[0112] Filtering and noise reduction: Using moving average or median filtering to suppress environmental interference noise;

[0113] Status self-check: If power supply abnormality, signal loss, or sensor failure is detected, the abnormal status will be reported immediately.

[0114] The sensor uploads data to the data acquisition edge in real time via wired (such as RS485 and CAN bus) or wireless (such as LoRa, WiFi and self-organizing network).

[0115] In this embodiment of the invention, the data fusion edge and storage layer assemble the raw data from the sensor, including the following steps:

[0116] A1. Determine the measurement range based on the communication protocols and raw data of each sensor;

[0117] A2. Perform self-tests according to the communication protocols of each sensor, retain the sensors with normal self-test status, and proceed to A3.

[0118] A3. Compare the sensor's raw data with its range, retain the sensors whose raw data is within the range, and proceed to A4.

[0119] A4. If the sensor is mounted on a bracket, assemble the sensor's motion information into JSON format data; otherwise, assemble the sensor's motion information into a dictionary.

[0120] For pressure / displacement sensors, a moving average filtering algorithm is used. The sensor's deviceId is retrieved from the dictionary above, along with the sensor data list, assuming a data length of [value missing]. The algorithm is as follows:

[0121] ;

[0122] in, Indicates within the time window The moving average value within the range is used to replace the original instantaneous sensor output to reduce the effects of fluctuations and noise. express The raw value collected by the corresponding sensor (deviceId) at each sampling time.

[0123] The above algorithm This represents the queue number of this sample. This represents the length of the queue.

[0124] For sensors such as gas and temperature sensors, a first-order low-pass filtering algorithm is used. Let the smoothing factor be... The previous filtering result was: The result for this test is: .

[0125] Data is smoothed using a recursive method, with historical and current values ​​merged according to fixed weights, and the coefficients... Determines the smoothing intensity and responds quickly to overall trends in gas concentration. It can be dynamically adjusted to adapt to different working conditions.

[0126] For other sensor data, a moving average filtering algorithm is used to address the interference from instantaneous fluctuations in the sensor data.

[0127] The pressure sensor data from the fully mechanized mining face needs to be combined with the support movement data and then processed in the following special way.

[0128] If the support is detected to rise from a stationary position, the working resistance and support movement data at this time should be recorded simultaneously. After the support rises, the support pressure data should be continuously observed. If the pressure tends to stabilize (the pressure change is less than the set threshold within 30 seconds), the pressure sensor data at this time should be recorded. The pressure sensor value at this time is the initial support force of the support.

[0129] The algorithm is as follows: from the moment the column rises is completed... From now on, continue to observe the pressure value sequence. The time interval must be at least 30 seconds. This indicates that after the column is raised, at time [time missing] The collected stent pressure values, This indicates that after the column is raised, at time [time missing] The collected stent pressure values, This indicates that after the column is raised, at time [time missing] The collected stent pressure value.

[0130] Set pressure stabilization threshold (e.g., 0.5 MPa). Calculate the difference between the maximum and minimum pressure values ​​within 30 seconds. Its expression is:

[0131] ;

[0132] like If the data at the end of the pressure value sequence is positive, then the initial support force of this support is positive. Otherwise, continue observing until the condition is met, and finally assemble the data into JSON data. The table sets a preset pressure stabilization threshold, which is used to determine the initial support force. This represents the sequence of pressure sensor data during the monitoring process of the support.

[0133] If the support is detected to move from a stationary position, the pressure sensor data value at that moment is recorded. The final resistance of the support is denoted as , and the data is ultimately assembled into JSON data.

[0134] For the fully mechanized mining face, in each sampling cycle, the current working resistance of all hydraulic supports is read and organized into a dictionary. Its expression is:

[0135] ;

[0136] Synchronously read the real-time advance rate (meters advanced) of the working face. Pre-configure the rated operating resistance F_rated of the pump station. In the middle, statistics satisfy Number of stents ,

[0137] Calculate the excess ratio :

[0138] ;

[0139] like Exceeding the basic alarm threshold If the pressure is detected, it is determined to be a "pressure incoming" state; otherwise, it is considered normal and no alarm is required. After determining "pressure incoming", according to... Intensity level is determined by comparing with segmented thresholds: if The intensity is set to 3 (intense pressure); otherwise if If the intensity is 2 (moderate pressure), then the intensity is 1 (light pressure).

[0140] The aforementioned data is promptly pushed to the data processing middleware, and the data is synchronized to the database according to the established strategy to solidify historical data.

[0141] In this embodiment of the invention, predicting the predicted step distance and final intensity level of the early warning layer includes the following steps:

[0142] B1. Obtain the time window based on the raw data after assembly. The internal pressure, initial support force, and final resistance, among which, Represents a timestamp. This indicates the length of time it takes to sample backwards from the current moment;

[0143] B2. Normalize the pressure, initial support force, and final resistance within the time window;

[0144] B3. Normalize the top-level off-board data;

[0145] B4. Based on the normalized initial support force and normalized pressure, calculate the correlation between the initial support force and historical pressure fluctuations.

[0146] B5. Based on the normalized final resistance and normalized pressure, calculate the correlation between the final resistance and historical pressure fluctuations.

[0147] B6. Calculate the correlation between the initial support force and the delamination of the top plate based on the normalized initial support force and the normalized delamination.

[0148] B7. Calculate the correlation between the final resistance and the roof delamination based on the normalized final resistance and the normalized delamination.

[0149] B8. Construct feature vectors based on the correlation between initial support force and historical pressure fluctuations, the correlation between final resistance and historical pressure fluctuations, the correlation between initial support force and roof delamination, and the correlation between final resistance and roof delamination.

[0150] B9. Obtain the hidden state based on the feature vector;

[0151] B10. Perform weighted fusion on the hidden states to obtain a weighted feature vector;

[0152] B11. Obtain the step size prediction value based on the weighted feature vector;

[0153] B12. Generate the final intensity level;

[0154] B13. Use the total loss function to train and adjust the step size prediction and the final intensity level.

[0155] The mine pressure prediction and early warning layer is the core layer of the system. This layer performs data normalization processing based on the data in the database, builds a prediction and early warning model, and predicts the pressure step distance and pressure intensity based on the prediction and early warning model. At the same time, it compares the real-time alarm data with the predicted data, performs unsupervised learning, and corrects the model to achieve the goal of accurate prediction and early warning.

[0156] In this embodiment of the invention, in B2, the expression for normalizing the pressure is:

[0157] ;

[0158] ;

[0159]

[0160] in, This represents the normalized pressure. Indicates the first The original value of the support pressure at all times. This represents the arithmetic mean of pressure within a time window. This represents the standard deviation of pressure values ​​within a time window. This represents the width of the sliding window used for normalization. Represents a timestamp. Indicates the first The raw pressure values ​​collected at any time;

[0161] In B2, the expression for normalizing the initial support force is:

[0162] ;

[0163] in, This represents the initial support force after normalization. Record the current pressure sensor value at the moment when the pressure stabilizes and is confirmed.

[0164] In B2, the expression for normalizing the final resistance is:

[0165] ;

[0166] in, This represents the final resistance after normalization. This indicates the pressure value when the support frame is moved.

[0167] In B3, the expression for normalizing the top-level off-board data is:

[0168] ;

[0169] ;

[0170] in, This indicates the separation of layers after normalization. Indicates the average value of the separation layer. Indicates recent Mean depth of point delamination, Indicates a point in time Data on top slab delamination at that time Indicates a relative time index. Indicates recent Standard deviation of the point separation depth.

[0171] In this embodiment of the invention, B4, the historical time series pressure deviation is... Deviation from the initial support force of the current support Multiply and average to measure the coupling between the initial support force and the overall historical pressure fluctuation pattern. When A large initial support force indicates that anomalies (high or low) in the initial support force are often accompanied by the synchronous, amplified, or offsetting effects of previous time-series pressures. The degree of correlation between the initial support force and historical pressure fluctuations... The expression is:

[0172] ;

[0173] in, This represents the initial support force after normalization. This indicates a deviation in historical time series pressure. Indicates the window length. Indicates a relative time index;

[0174] In B5, the resistance deviation is measured when the support frame relocation is completed. The coupling relationship with historical pressure sequences. If a historical pressure fluctuation and the final resistance simultaneously increase or decrease, the crossover value rises, indicating the potential impact of the final resistance on subsequent pressure trends. The degree of correlation between the final resistance and historical pressure fluctuations. The expression is:

[0175] ;

[0176] in, This represents the final resistance after normalization;

[0177] In B6, a simple product captures the combined effect of the initial support force deviation and the average deviation of the roof delamination. If both deviate in the same direction (e.g., the initial support force is too high and the delamination is deep), the value is positive and has a large absolute value, indicating that the pressure and roof condition are both trending towards anomalies under this condition. The degree of correlation between the initial support force and the roof delamination. The expression is:

[0178] ;

[0179] In B7, the interaction between the final resistance deviation and the average deviation of roof delamination is captured. If the final resistance immediately after the support relocation is completed is accompanied by a large delamination, it indicates insufficient support control over the roof, potentially foreshadowing a higher accident risk. The degree of correlation between final resistance and roof delamination is discussed. The expression is:

[0180] ;

[0181] in, This indicates the separation of layers after normalization.

[0182] In this embodiment of the invention, in B8, the feature vector The expression is:

[0183] ;

[0184] in, Table Normalize stress values ​​at all times. Table Normalize stress values ​​at all times. Indicates in Normalize stress values ​​at all times. Represents a sequence of pressure increments. Represents a binary sequence of stent movements. Indicates real-time progress. Indicates geological structure parameters, This represents the initial support force after normalization. This represents the final resistance after normalization. This indicates the separation of layers after normalization. Indicates in The increment of delamination at the top slab at any given time. Indicates the pressure gradient between adjacent supports. Indicates the degree of coupling between pressure and action. Indicates the degree of coupling between pressure and propulsion. Indicates in The increment of delamination at the top slab at any given time. This indicates the degree of correlation between initial support force and historical pressure fluctuations. This indicates the degree of correlation between the final resistance and historical pressure fluctuations. This indicates the degree of correlation between the initial support force and the delamination of the top slab. This indicates the degree of correlation between the final resistance and the top plate delamination.

[0185] In this embodiment of the invention, B9 is a hidden state. The expression is:

[0186] ;

[0187] in, This represents a gated loop network. Represents the eigenvector. The parameters represent the gated recurrent network. Represents a real matrix. Indicates a time step. This indicates the length of the output vector at each time point;

[0188] In B10, the weighted eigenvectors The expression is:

[0189] ;

[0190] ;

[0191] in, This indicates the transpose operation. Represents the attention weight vector. This represents the activation function. Indicates attention weight;

[0192] In B11, the step size prediction value The expression is:

[0193] ;

[0194] in, Represents the regression weight matrix. Indicates the regression bias term;

[0195] In this embodiment of the invention, B12 includes the following sub-steps:

[0196] B121. Based on the normalized pressure, the multi-channel time series characteristic sequence is obtained;

[0197] B122. Input the multi-channel time-series feature sequence into a bidirectional LSTM;

[0198] B123. Based on the state vector output by the bidirectional LSTM, obtain the probability value for each intensity level;

[0199] B124. Based on the probability value of each intensity level, the final intensity level is obtained.

[0200] In this embodiment of the invention, in B121, firstly, a one-dimensional convolutional neural network is applied to the standardized pressure sequence. One-dimensional convolution addresses the signal characteristics of high-intensity pressure events, which are often accompanied by abrupt changes and violent fluctuations. Compared to traditional statistical methods, it more effectively captures the change patterns within these local time windows. Multi-channel time-series feature sequences. The expression is:

[0201] ;

[0202] in, Represents the normalized pressure sequence. Indicates the kernel size. Indicates the number of convolution kernels. Indicates the step size. This represents a one-dimensional convolutional neural network;

[0203] In B122, a bidirectional LSTM (BiLSTM) is introduced to model the contextual information throughout the stress evolution process. The state vector output by the bidirectional LSTM... The expression is:

[0204] ;

[0205] in, This represents the hidden unit dimension of a bidirectional LSTM. Indicates a bidirectional LSTM;

[0206] In B123, the probability value for each intensity level. The expression is:

[0207] ;

[0208] in, This represents the activation function. Indicates the first Input data Representation vector, Represents a linear weight matrix. Indicates the bias term;

[0209] In B124, the final intensity level The expression is:

[0210] ;

[0211] in, This represents the set of independent variables that cause the function to reach its maximum value;

[0212] In B13, the total loss function The expression is:

[0213] ;

[0214] ;

[0215] ;

[0216] in, The weights represent the regression task loss. The weights represent the loss of the classification task. This indicates the loss of the return mission. Indicates the loss of the classification task. Indicates the first The step size predicted by the sample model. Indicates the first The actual step distance of each sample. Indicates the first The coding of the intensity label of the sample, The model predicts the first... The sample belongs to the intensity category The probability, Represents the total number of samples. Represents a logarithmic function.

[0217] In this embodiment of the invention, the above-mentioned algorithm and dataset are used to train the model. At the same time, the pressure step distance and pressure intensity in the historical dataset are compared and analyzed with the predicted pressure intensity and pressure step distance in the training dataset. The Adam optimizer is introduced to perform an adaptive learning rate algorithm to perform unsupervised adaptive correction and prevent the model from overfitting.

[0218] To address model drift caused by changes in mining pressure environment, the system continuously collects feedback data after deployment, triggering online fine-tuning:

[0219] Error detection and real-time calculation of the deviation between the prediction and the actual value. and intensity prediction error indication Their expressions are as follows:

[0220] ;

[0221] ;

[0222] in, This indicates the current actual step size. Indicates the true intensity. This indicates the final intensity level.

[0223] If any condition is met within the most recent sliding window, adaptive fine-tuning is initiated:

[0224] For example, the prediction step error exceeds the tolerance threshold. (e.g., 0.5 meters).

[0225] For example: the number of intensity classification errors accumulates to more than [number missing] within the window. times (e.g., 5 times).

[0226] The model fine-tuning strategy is as follows:

[0227] Except for freezing the attention layer and output layer, the other parameters (GRU / LSTM / convolutional layer weights) remain unchanged to ensure the stability of the model backbone.

[0228] Train on incremental data and collect the most recent For each trigger sample, fine-tune it in small batches, updating only a few of the above parameters.

[0229] After fine-tuning, the new parameters smoothly replace the online model without interrupting the prediction service.

[0230] Through the training and verification of the model, the predicted pressure step distance and pressure intensity of the current fully mechanized mining face are continuously output. If the prediction results reach the set early warning conditions, the data is automatically pushed to the application layer.

[0231] As the outermost layer of the system, the application layer continuously receives pressure step distance and intensity results pushed by the prediction and early warning layer. After parsing, it displays these results intuitively on the monitoring screen and mobile devices using charts, curves, and highlighted warnings. When the step distance falls below the safety threshold or the intensity reaches the "severe" level, it automatically triggers audible and visual alarms, SMS / WeChat, and email notifications. Users can also switch to historical data viewing with one click, providing real-time decision support for dispatching and command and on-site operations. This layer can also be linked with hydraulic pump stations and tunneling equipment to automatically issue pressure limiting, pressure reducing, or deceleration commands, effectively transforming early warnings into safety control measures.

[0232] Meanwhile, the application layer records all alarm events, on-site feedback, and control commands, stores them in a log database, and regularly generates accuracy assessment reports, supporting online adaptive model correction and safety management audits. After on-site personnel confirm the actual time and intensity of the pressure on their mobile devices, the system automatically sends the feedback data back to the prediction and early warning layer, realizing a closed loop of "early warning → alarm → control → feedback," continuously optimizing model performance and minimizing mine pressure risks.

[0233] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader 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. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

[0234] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader 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. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A data processing based environment warning system, characterized by, The system comprises a perception layer, a data fusion edge and storage layer, a prediction and early warning layer, and an application layer; The perception layer is configured to collect raw data through a plurality of sensors; The data fusion edge and storage layer is configured to assemble and process the raw data of the sensors; The prediction and early warning layer is configured to predict a step distance prediction value and a final intensity level according to the assembled and processed raw data; The application layer is configured to push the step distance prediction value and the final intensity level; According to the layout of the mine roadway, the fully mechanized working face, the key support structure and the geological conditions, various sensors are arranged, including support pressure sensors, travel displacement sensors, roof separation sensors and environmental sensors; The prediction and early warning layer predicting the step distance prediction value and the final intensity level comprises the following steps: B1, obtaining a time window according to the original data after the assembly processing the pressure in the inner portion, the initial supporting force and the final resisting force, wherein, represents a time stamp, represents a time length of sampling back from the current time B2, normalizing the pressure, the initial support force and the final resistance in the time window; B3, normalizing the roof separation data; B4, calculating the correlation degree of the initial support force and the historical pressure fluctuation according to the normalized initial support force and the normalized pressure; B5, calculating the correlation degree of the final resistance and the historical pressure fluctuation according to the normalized final resistance and the normalized pressure; B6, calculating the correlation degree of the initial support force and the roof separation according to the normalized initial support force and the normalized separation; B7, calculating the correlation degree of the final resistance and the roof separation according to the normalized final resistance and the normalized separation; B8, constructing a feature vector according to the correlation degree of the initial support force and the historical pressure fluctuation, the correlation degree of the final resistance and the historical pressure fluctuation, the correlation degree of the initial support force and the roof separation, and the correlation degree of the final resistance and the roof separation; B9, obtaining a hidden state according to the feature vector; B10, obtaining a weighted feature vector by weighted fusion of the hidden state; B11, obtaining a step distance prediction value according to the weighted feature vector; B12, generating a final intensity level; B13, training and adjusting the step distance prediction value and the final intensity level by using a total loss function; In the B11, the step prediction value The expression is: ； wherein, represents a regression weight matrix, represents a regression bias term; The B12 comprises the following sub-steps: B121, obtaining a multi-channel time sequence feature sequence according to the normalized pressure; B122, inputting the multi-channel time sequence feature sequence into a bidirectional LSTM; B123, obtaining a probability value of each intensity level according to a state vector output by the bidirectional LSTM; B124, obtaining a final intensity level according to the probability value of each intensity level; In the B13, the total loss function The expression is: ； ； ； wherein, represents a weight of the regression task loss, represents a weight of the classification task loss, represents the regression task loss, represents the classification task loss, represents the predicted step distance of the i-th sample model, represents the predicted step distance of the i-th sample model, represents the true step distance of the i-th sample, represents the true step distance of the i-th sample, represents the encoding of the intensity label of the i-th sample, represents the encoding of the intensity label of the i-th sample, represents the probability that the i-th sample belongs to the intensity category represents the probability that the i-th sample belongs to the intensity category represents the probability that the i-th sample belongs to the intensity category represents the total number of samples, represents the logarithm function.

2. The data processing based environmental warning system of claim 1, wherein, The data fusion edge and storage layer assembling and processing the raw data of the sensors comprises the following steps: A1, determining a range according to the communication protocol and the raw data of each sensor; A2, performing self-checking according to the communication protocol of each sensor, retaining the sensors with normal self-checking states, and entering A3; A3, comparing the raw data of the sensors with the range, retaining the sensors with the raw data within the range, and entering A4; A4, if the sensor is installed on a support, assembling action information of the sensor to generate json format data, otherwise assembling the action information of the sensor to generate a dictionary.

3. The data processing based environmental warning system of claim 1, wherein, In the B2, the expression for normalizing the pressure is: ； ； wherein, represents the normalized pressure, represents the raw pressure value collected at the moment of time, represents the arithmetic mean of the pressure within the time window, represents the standard deviation of the pressure values within the time window, represents the sliding window width used for normalization, represents the time stamp, represents the raw pressure value collected at the moment of time; In the B2, the expression for normalizing the initial support force is: ； wherein, represents the normalized initial support force, represents recording the current pressure sensor value at the time when the pressure is stabilized In the B2, the expression for normalizing the final resistance is: ； wherein, represents the normalized end resistance, represents the pressure value at the end of the support shift; In the B3, the expression for normalizing the top layer off-board data is: ； ； in, This indicates the separation of layers after normalization. Indicates the average value of the separation layer. Indicates recent Mean depth of point delamination, Indicates a point in time Data on top slab delamination at that time Indicates a relative time index. Indicates recent Standard deviation of the point separation depth.

4. The data processing based environmental warning system of claim 1, wherein, In the B4, the degree of correlation between the initial support force and the historical pressure fluctuation The expression is: ； wherein, represents the normalized initial support force, represents the historical time series pressure deviation, represents the window length, represents the relative time index; In the B5, the degree of correlation between the resistance and the historical pressure fluctuation The expression is: ； wherein, represents the normalized end resistance; The B6, the initial support force and the roof separation layer associated degree The expression is: ； The correlation degree between the final resistance and the roof separation layer in the B7 The expression is: ； wherein, represents the normalized delamination.

5. The data processing based environmental warning system of claim 1, wherein, In the B8, the eigenvector The expression is: ； wherein, Table 1 normalized pressure value at time t, Table 1 normalized pressure value at time t, normalized pressure value at time t, normalized pressure value at time t, pressure increment sequence, binary sequence of support actions, real-time advancement, geological structure parameters, normalized initial support force, normalized final resistance force, normalized separation, roof separation increment at time t, roof separation increment at time t, adjacent support pressure gradient, pressure-action coupling degree, pressure-advancement coupling degree, roof separation increment at time t, roof separation increment at time t, initial support force-historical pressure fluctuation correlation degree, final resistance force-historical pressure fluctuation correlation degree, initial support force-roof separation correlation degree, final resistance force-roof separation correlation degree.

6. The data processing based environmental warning system of claim 1, wherein, In the B9, the hidden state The expression is: ； wherein, denotes a gated recurrent network, denotes a feature vector, denotes a parameter of the gated recurrent network, denotes a real matrix, denotes a window length, denotes a length of an output vector at each time point; In the B10, the weighted feature vector The expression is: ； ； wherein, denotes a transpose operation, denotes an attention weight vector, denotes an activation function, denotes an attention weight.

7. The data processing based environmental warning system of claim 1, wherein, In the B121, the multi-channel time sequence feature sequence The expression is: ； wherein, denotes a normalized pressure sequence, denotes a convolution kernel size, denotes a number of convolution kernels, denotes a step size, denotes a one-dimensional convolutional neural network; In the B122, the state vector output by the bidirectional LSTM The expression is: ； wherein, represents the dimension of the hidden units of the bidirectional LSTM, represents the bidirectional LSTM; In the B123, the probability value of each intensity level The expression is: ； wherein, represents an activation function, represents the first input data, characterizing vector, represents a linear weight matrix, represents a bias term; In the B124, the final intensity level is expressed by the formula: ； wherein denotes the set of arguments for which the function attains its maximum value.

Images