Intelligent monitoring method for underground coal mine equipment based on Internet of Things

Abstract

The invention relates to the technical field of coal mine underground equipment intelligent monitoring, in particular to a coal mine underground equipment intelligent monitoring method based on the Internet of Things. Comprising the following steps: collecting equipment vibration, temperature and current data through multiple types of sensors; adaptive Ka lman filtering is carried out at edge nodes to suppress noise, and multi-modal fusion features (FFT vibration spectrum, temperature root mean square and current peak) are extracted; the preprocessed data are uploaded to a central server, a deep reinforcement learning model (including an LSTM network and a strategy network) is adopted to analyze the equipment state, a fault is predicted, and a control signal is generated; dynamically adjusting the weight of the DRL model and the sampling frequency of the sensor based on the real-time environmental parameters; the data integrity is verified through a block chain mechanism, and the communication efficiency between devices is optimized by using collaborative learning. According to the method, the problems of high hysteresis, poor environmental adaptability and insufficient utilization of multi-source data of a traditional monitoring technology are solved, and the fault early warning accuracy and the system robustness are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for underground equipment in coal mines, and specifically to an intelligent monitoring method for underground equipment in coal mines based on the Internet of Things. Background Technology

[0002] The underground working environment in coal mines is characterized by high risk, strong concealment, and complexity and variability. Equipment operates under conditions of high load, high temperature, high humidity, and high dust for extended periods, which can easily lead to safety accidents such as mechanical failures and electrical short circuits. Traditional equipment monitoring technologies mainly rely on threshold alarm mechanisms and periodic manual inspections, which have significant shortcomings.

[0003] The monitoring lag is significant: threshold alarms are only triggered when equipment parameters exceed preset ranges, failing to provide early warning of faults. For example, progressive faults such as bearing wear and motor insulation aging often result in irreversible damage or even trigger a chain reaction of accidents by the time temperature or vibration exceeds limits.

[0004] Severe environmental interference: Electromagnetic interference, mechanical vibration, and humidity fluctuations in the well result in low signal-to-noise ratios of sensor data. Conventional filtering algorithms (such as moving average methods) struggle to adapt to nonlinear noise, leading to a surge in false alarm rates; fixed sampling frequencies cause a sharp drop in data validity in high-humidity environments, affecting the accuracy of condition assessments.

[0005] Fragmented multi-source data: Existing systems typically process single-mode data such as vibration, temperature, and current independently, lacking feature-level fusion mechanisms. For example, relying solely on peak current may misjudge instantaneous overload as a fault, while ignoring the correlation analysis between vibration spectrum and temperature rise trend leads to one-sided diagnosis.

[0006] The system lacks adaptability: Most models use static parameter configurations and cannot respond to dynamic changes in the underground environment (such as sudden changes in temperature and humidity caused by the advancement of the mining face). Although deep learning models have certain predictive capabilities, they do not introduce real-time environmental feedback mechanisms, and the generalization ability of the models degrades as the working conditions change.

[0007] Data security is weak: data transmission from edge nodes to servers is vulnerable to malicious tampering, and traditional encryption protocols are inefficient on resource-constrained mining equipment, making it difficult to balance the needs of real-time performance and integrity verification.

[0008] Therefore, there is an urgent need to build a monitoring system that integrates edge intelligence, multimodal collaborative analysis, and environmental adaptation to overcome the technical bottlenecks of low accuracy in downhole equipment status perception, delayed fault prediction, and insufficient system robustness.

[0009] Therefore, the existing technology still needs further development. Summary of the Invention

[0010] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide an intelligent monitoring method for underground equipment in coal mines based on the Internet of Things, so as to solve the problems existing in the prior art.

[0011] To achieve the above-mentioned technical objectives, according to a first aspect of the present invention, the present invention provides an intelligent monitoring method for underground equipment in coal mines based on the Internet of Things, comprising:

[0012] S100. Collect status data of the target device through multiple IoT sensors deployed underground in the coal mine. The status data includes vibration data, temperature data, and current data.

[0013] S200. The edge computing nodes set up downhole preprocess the collected data, including noise suppression and feature extraction; the preprocessed data is then transmitted to the central server via an Internet of Things (IoT) network.

[0014] S300: The central server uses a deep reinforcement learning model to analyze preprocessed data to assess equipment status and predict potential failures; based on the prediction results, alarms or control signals are generated and output to the execution device; the decision parameters of the deep reinforcement learning model are adjusted according to real-time environmental feedback to achieve adaptive optimization.

[0015] Specifically, the noise suppression step includes using an adaptive Kalman filter, whose update formula is:

[0016]

[0017] in:

[0018] The state estimate at the current time step represents the filtered sensor data value.

[0019] The state estimate of the previous time step represents the filtered output of the previous moment;

[0020] K k Kalman gain matrix, representing adaptively adjusted weights used for prediction and measurement based on noise statistical balance;

[0021] z k The measured value at the current time step represents the raw sensor data.

[0022] H: Observation matrix, representing the transformation coefficients from state to measurement.

[0023] Specifically, the feature extraction step includes calculating multimodal time series features, the formula of which is:

[0024] f(t) = α·FFT(v t)+β·RMS(T t )+γ·I peak (t)

[0025] in:

[0026] f(t): The value of the extracted feature vector at time t, representing the feature representation of the fusion of vibration, temperature and current;

[0027] α, β, γ: Weighting coefficients, representing the relative importance factors of each sensor's data;

[0028] FFT(v t ): The Fast Fourier Transform (FFT) value of the vibration data at time t, representing the vibration frequency characteristics; RMS(T) t ): The root mean square value of temperature data at time t, representing the characteristics of temperature fluctuation;

[0029] I peak (t): The peak value of the current data at time t, representing the characteristics of current overload risk.

[0030] Specifically, the deep reinforcement learning model includes an LSTM network and a policy network, wherein the state output formula of the LSTM network is:

[0031] h t =LSTM(x t ,h t-1 ,c t-1 )

[0032] in:

[0033] h t The hidden state at the current time step represents the time-series characteristics of the device state.

[0034] x t Input data, representing the preprocessed feature vectors;

[0035] h t-1 The hidden state of the previous time step represents the memory of the historical state;

[0036] c t-1 The cell state at the previous time step represents long-term dependency information.

[0037] Specifically, the method also includes using blockchain security mechanisms to verify data transmission, specifically:

[0038] Generate data hashes for key events at edge nodes;

[0039] Record the hash value to a lightweight blockchain network;

[0040] The hash generation formula is:

[0041] Hash=SHA-256(data+timestamp)

[0042] in:

[0043] Hash: Outputs a hash value, representing a unique and immutable identifier for the data;

[0044] data: Input data, representing compressed sensor data packets;

[0045] timestamp: A timestamp that represents the absolute time of data collection.

[0046] Specifically, the adaptive optimization step includes dynamically adjusting the sensor sampling frequency based on the ambient humidity parameter, and the adjustment formula is as follows:

[0047] f samp =f min +(f max -f min )·σ(Δh)

[0048] in:

[0049] f samp The adjusted sampling frequency represents the sensor's data acquisition rate.

[0050] f min Minimum allowed sampling frequency, representing the preset base sampling rate;

[0051] f max : Maximum allowed sampling frequency, representing the preset maximum sampling rate;

[0052] σ(Δh): Sigmoid activation function, representing a modulating factor mapped according to the rate of change of humidity Δh (representing the derivative of ambient humidity).

[0053] Specifically, the policy network outputs an action-value function, the formula of which is:

[0054]

[0055] in:

[0056] Q(s,a): State-action value, representing the expected long-term reward of taking action a in state s;

[0057] s: Current state, representing the device state characteristics output by the LSTM;

[0058] a: Current action, representing a control decision such as "initiating frequency reduction" or "sending an alarm";

[0059] r: Instant reward, representing feedback calculated based on prediction accuracy;

[0060] γ: Discount factor, representing the weight of future returns (0 < γ < 1);

[0061] s′: Next state, representing the state predicted by the LSTM at the next time step.

[0062] Specifically, the method also includes inter-device collaborative learning, wherein the loss function is:

[0063] L=λ1L pred +λ2L comm

[0064] in:

[0065] L: Total loss value, representing the model optimization objective;

[0066] L pred : Predicted loss, representing the error in fault prediction;

[0067] L comm Communication loss represents the overhead of data transmission between devices;

[0068] λ1, λ2: Weighting parameters, representing the balance between prediction accuracy and communication cost.

[0069] Specifically, the alarm triggering step includes calculating a risk index, the formula of which is:

[0070]

[0071] in:

[0072] R: Risk Index, representing the overall risk score of equipment failure;

[0073] P fail Fault probability, representing the output value of the deep reinforcement learning model;

[0074] I severity Severity coefficient: Represents the degree of harm that the fault may cause;

[0075] E energy Energy consumption value, representing the current power consumption of the device.

[0076] Specifically, the adaptive optimization step includes adjusting the DRL model parameters using an environmental adaptive factor, the formula of which is:

[0077] θ new =θ old ·η·e -α·env_var

[0078] in:

[0079] θ newThe updated model parameters represent the adjusted weights.

[0080] θ old Original model parameters, representing the initial weights;

[0081] η: Learning rate, representing the update step size factor;

[0082] α: Attenuation coefficient, representing the sensitivity to environmental influences (0 < α < 1);

[0083] env_var: An environment variable representing the real-time monitored humidity or temperature value.

[0084] Beneficial effects:

[0085] This invention significantly improves the intelligence level and reliability of underground equipment monitoring systems in coal mines through the integration and innovation of multiple technologies.

[0086] 1. Achieve accurate early warning of faults:

[0087] Deep reinforcement learning models, combined with LSTM time series analysis and policy decision-making, effectively capture the progressive degradation characteristics of equipment, overcoming the lag bottleneck of traditional threshold alarms. A multi-modal feature fusion mechanism collaboratively analyzes vibration spectra, temperature fluctuations, and current peaks, eliminating the risk of misjudgment from single-source data and improving the ability to identify complex faults.

[0088] 2. Enhance adaptability to complex environments:

[0089] The sensor sampling frequency is dynamically adjusted based on the humidity change rate, and a smooth transition is achieved through the Sigmoid function to ensure the validity of data under high humidity conditions. Environmental adaptive factors are incorporated into the DRL model parameters for real-time optimization, suppressing model performance degradation caused by sudden temperature and humidity changes and maintaining prediction stability.

[0090] 3. Build a trusted data chain:

[0091] A lightweight blockchain mechanism generates immutable data hashes, and combined with an efficient consensus algorithm, it verifies the transmission integrity from edge nodes to servers, overcoming the security bottleneck of the Industrial Internet of Things in resource-constrained environments.

[0092] 4. Optimize system resource efficiency:

[0093] Edge computing nodes perform adaptive Kalman filtering and feature extraction, significantly reducing data transmission volume. A collaborative learning mechanism balances inter-device communication overhead and prediction accuracy, reducing redundant communication load. An adaptive sampling strategy reduces energy consumption from ineffective data acquisition on high-load devices.

[0094] 5. Supports precise risk management:

[0095] The risk index model quantifies the probability of failure, the degree of harm, and equipment energy consumption, driving a graded response mechanism (shutdown / frequency reduction / alarm) to achieve precise scheduling of maintenance resources and intelligent decision-making for safe production. Attached Figure Description

[0096] Figure 1 This is a flowchart illustrating the intelligent monitoring method for underground equipment in coal mines based on the Internet of Things, provided in a specific embodiment of the present invention.

[0097] Figure 2 This is a schematic diagram of the system composition of an IoT-based intelligent monitoring system for underground equipment in coal mines, provided in a specific embodiment of the present invention. Detailed Implementation

[0098] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments in this application, other similar embodiments obtained by those skilled in the art without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0099] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0100] Please see Figure 1 This invention provides an intelligent monitoring method for underground equipment in coal mines based on the Internet of Things, comprising:

[0101] S100. Collect status data of the target device by multiple IoT sensors deployed underground in the coal mine. The status data includes vibration data, temperature data and current data.

[0102] It should be further noted that the system architecture and hardware configuration designed in this invention include:

[0103] 1. Sensor layer:

[0104] Deploy in critical parts of the equipment:

[0105] Vibration sensor (preferably IEPE type, range ±50g, default sampling rate 200Hz);

[0106] Temperature sensor (PT1000 platinum resistance thermometer, accuracy ±0.5℃);

[0107] Current sensor (Hall effect type, range 0-200A);

[0108] Reason: High range and accuracy ensure data reliability under high load conditions in coal mines.

[0109] 2. Edge computing layer:

[0110] An embedded node using an ARM Cortex-A72 core (4GB memory) is deployed as follows:

[0111] Adaptive Kalman filter;

[0112] Feature extraction module;

[0113] Lightweight blockchain client.

[0114] 3. Central Server:

[0115] Configured with an Intel Xeon Gold 6338 processor, running a deep reinforcement learning (DRL) model.

[0116] S200. The edge computing nodes set up downhole preprocess the collected data, including noise suppression and feature extraction; the preprocessed data is then transmitted to the central server via an Internet of Things network.

[0117] Specifically, the noise suppression step includes using an adaptive Kalman filter, whose update formula is:

[0118]

[0119] in:

[0120] The state estimate at the current time step represents the filtered sensor data value.

[0121] The state estimate of the previous time step represents the filtered output of the previous moment;

[0122] K k Kalman gain matrix, representing adaptively adjusted weights used for prediction and measurement based on noise statistical balance;

[0123] z k The measured value at the current time step represents the raw sensor data.

[0124] H: Observation matrix, representing the transformation coefficients from state to measurement;

[0125] P k|k-1 State prediction covariance, physically representing the confidence level of the estimation error.

[0126] R k: Measure noise covariance, preferred value: diag(0.01,0.02,0.05) (corresponding to vibration, temperature and current respectively), reason: low noise covariance setting can adapt to signal fluctuations caused by downhole dust.

[0127] Specifically, the feature extraction step includes calculating multimodal time series features, the formula of which is:

[0128] f(t) = α·FFT(v t )+β·RMS(T t )+γ·I peak (t)

[0129] in:

[0130] f(t): The value of the extracted feature vector at time t, representing the feature representation of the fusion of vibration, temperature and current;

[0131] α, β, γ: Weighting coefficients, representing the relative importance factors of each sensor's data;

[0132] FFT(v t ): The fast Fourier transform value of vibration data at time t, representing the vibration frequency characteristics;

[0133] RMS(T t ): The root mean square value of temperature data at time t, representing the characteristics of temperature fluctuation;

[0134] I peak (t): The peak value of the current data at time t, representing the characteristics of current overload risk.

[0135] It should be further explained that, regarding the feature fusion formula, the present invention is further designed as follows:

[0136]

[0137] in:

[0138] FFT(v t [5-100Hz]: Energy sum of the vibration spectrum in the 5-100Hz range (representing the characteristic frequency band of bearing failure);

[0139] Temperature change rate (representing an overheating risk indicator);

[0140] I crest (t): Current peak factor (representing the characteristics of motor insulation failure);

[0141] Preferred weight values: α = 0.6, β = 0.3, γ = 0.1. Reason: Vibration data contributes the most to mechanical fault diagnosis, and experiments show that the F1-score of this weight combination reaches 92%.

[0142] Specifically, the method also includes using blockchain security mechanisms to verify data transmission, specifically:

[0143] Generate data hashes for key events at edge nodes;

[0144] Record the hash value to a lightweight blockchain network;

[0145] The hash generation formula is:

[0146] Hash=SHA-256(data+timestamp)

[0147] in:

[0148] Hash: Outputs a hash value, representing a unique and immutable identifier for the data;

[0149] data: Input data, representing compressed sensor data packets;

[0150] timestamp: A timestamp that represents the absolute time of data collection.

[0151] It should be further explained that, regarding the blockchain security mechanism, this invention has made the following further design:

[0152] Step 1: Hash Generation

[0153] formula:

[0154] Hash i =SHA-256(f(t) i )||timestamp i )

[0155] in:

[0156] ||: Data concatenation operation, representing the binding of feature values ​​and time;

[0157] Timestamp precision: 1ms, to prevent replay attacks.

[0158] 2. Consensus Mechanism in Step 2:

[0159] Using a simplified BFT algorithm:

[0160] The verification group consists of 4 edge nodes;

[0161] After 3 / 4 of the nodes reach a consensus, the data is written into the block;

[0162] Block generation interval: 2 seconds;

[0163] Reason: Short intervals meet the real-time requirements of downhole drilling, with an experimental delay of <500ms.

[0164] S300: The central server uses a deep reinforcement learning model to analyze preprocessed data to assess equipment status and predict potential failures; based on the prediction results, alarms or control signals are generated and output to the execution device; the decision parameters of the deep reinforcement learning model are adjusted according to real-time environmental feedback to achieve adaptive optimization.

[0165] Specifically, the deep reinforcement learning model includes an LSTM network and a policy network, wherein the state output formula of the LSTM network is:

[0166] h t =LSTM(x t ,h t-1 ,c t-1 )

[0167] in:

[0168] h t The hidden state at the current time step represents the time-series characteristics of the device state.

[0169] x t Input data, representing the preprocessed feature vectors;

[0170] h t-1 The hidden state of the previous time step represents the memory of the historical state;

[0171] c t-1 The cell state at the previous time step represents long-term dependency information.

[0172] Specifically, the adaptive optimization step includes dynamically adjusting the sensor sampling frequency based on the ambient humidity parameter, and the adjustment formula is as follows:

[0173] f samp =f min +(f max -f min )·σ(Δh)

[0174] in:

[0175] f samp The adjusted sampling frequency represents the sensor's data acquisition rate.

[0176] f min Minimum allowed sampling frequency, representing the preset base sampling rate;

[0177] f max: Maximum allowed sampling frequency, representing the preset maximum sampling rate;

[0178] σ(Δh): Sigmoid activation function, representing a modulating factor mapped according to the rate of change of humidity Δh (representing the derivative of ambient humidity).

[0179] It should be further noted that the method includes designing a frequency adjustment formula:

[0180]

[0181] in:

[0182] Δh: Humidity change rate (%) / min, representing the increase in sampling rate when humidity rises sharply;

[0183] Operating parameters:

[0184] f min =50Hz (basic monitoring);

[0185] f max =200Hz (high-precision monitoring);

[0186] Reason for optimal parameter selection: The Sigmoid function provides a smooth transition, avoiding mechanical shock interference.

[0187] Specifically, the policy network outputs an action-value function, the formula of which is:

[0188]

[0189] in:

[0190] Q(s,a): State-action value, representing the expected long-term reward of taking action a in state s;

[0191] s: Current state, representing the device state characteristics output by the LSTM;

[0192] a: Current action, representing a control decision such as "initiating frequency reduction" or "sending an alarm";

[0193] r: Instant reward, representing feedback calculated based on prediction accuracy;

[0194] γ: Discount factor, representing the weight of future returns (0 < γ < 1);

[0195] s′: Next state, representing the state predicted by the LSTM at the next time step.

[0196] It should be further explained that, regarding the deep reinforcement learning model, this invention has made the following further design:

[0197] Step 1 Model Structure:

[0198] LSTM layer: 128 neurons, time step T = 60 (representing a 1-minute data window);

[0199] Policy network: 3 fully connected layers (64-32-16 neurons), output action space: {downtime, shutdown, alarm, maintain};

[0200] Step 2: DRL Training (Central Server)

[0201] Reward function formula:

[0202]

[0203] in:

[0204] λ: Energy consumption factor (preferred value: 0.3), representing the balance coefficient between energy saving and safety;

[0205] E saved Energy saved during action execution (kW·h);

[0206] Reason: By penalizing false positives and false negatives, the model reduced the false positive rate to 6.8% on the test set;

[0207] Step 2.3 Adaptive parameter update:

[0208] Environmental drive parameter adjustment formula:

[0209]

[0210] in:

[0211] h: Real-time humidity (%), representing the reduction in model update magnitude due to high humidity;

[0212] η: Learning rate (preferred value: 0.001), representing the step size of weight updates. Reason: The exponential decay term prevents the model from overfitting in high humidity environments.

[0213] θ DRL The parameter set (such as neural network weights) of a deep reinforcement learning (DRL) model represents the internal parameters of the model that need to be dynamically adjusted, which directly affects the accuracy of fault prediction.

[0214] The gradient of the loss function J(θ) with respect to the parameter θ indicates the direction of parameter adjustment, optimizing the model towards reducing prediction error. Actual calculation: based on the difference between the predicted result and the actual fault label (preferably the mean squared error in this invention).

[0215] Specifically, the method also includes inter-device collaborative learning, wherein the loss function is:

[0216] L=λ1Lpred +λ2L comm

[0217] in:

[0218] L: Total loss value, representing the model optimization objective;

[0219] L pred : Predicted loss, representing the error in fault prediction;

[0220] L comm Communication loss represents the overhead of data transmission between devices;

[0221] λ1, λ2: Weighting parameters, representing the balance between prediction accuracy and communication cost.

[0222] Specifically, the alarm triggering step includes calculating a risk index, the formula of which is:

[0223]

[0224] in:

[0225] R: Risk Index, representing the overall risk score of equipment failure;

[0226] P fail Fault probability, representing the output value of the deep reinforcement learning model;

[0227] I severity Severity coefficient: Represents the degree of harm that the fault may cause;

[0228] E energy Energy consumption value, representing the current power consumption of the device.

[0229] Specifically, the adaptive optimization step includes adjusting the DRL model parameters using an environmental adaptive factor, the formula of which is:

[0230] θ new =θ old ·η·e -α·env_var

[0231] in:

[0232] θ new The updated model parameters represent the adjusted weights.

[0233] θ old Original model parameters, representing the initial weights;

[0234] η: Learning rate, representing the update step size factor;

[0235] α: Attenuation coefficient, representing the sensitivity to environmental influences (0 < α < 1);

[0236] env_var: An environment variable representing the real-time monitored humidity or temperature value.

[0237] It should be further explained that the present invention specifically designs a collaborative learning and risk decision-making scheme, including:

[0238] 1. Step 1: Knowledge Sharing Between Devices

[0239] Loss function:

[0240]

[0241] in:

[0242] L: The total loss function value for collaborative learning, representing the combined optimization objective of prediction accuracy and communication overhead. This value needs to be minimized to balance model performance and system resource consumption.

[0243] Mean Square Error (MSE) of Equipment Condition Prediction, Component Description:

[0244] N: Number of training samples (preferred value: 200), which physically means the data packet capacity of a single batch. This value takes into account both the computing power of edge nodes and the real-time requirements.

[0245] y i The actual status label of the equipment (0: normal, 1: fault) is derived from the historical database of sensor and manual diagnostic records;

[0246] The model-predicted failure probability values ​​(in the range [0,1]) are output by the DRL network.

[0247] Inter-device communication constraints penalize excessive differences in model parameters between adjacent devices to control energy consumption. Component descriptions:

[0248] Adjacent device groups are physically defined as groups of devices with a signal transmission distance ≤ 50 meters. This threshold ensures the stability of wireless communication (determined by actual measurements using an underground roadway attenuation model).

[0249] θ j : The local DRL model parameter vector of the j-th device, representing the device-specific fault diagnosis knowledge;

[0250] θ avg Average parameters of neighboring devices, calculation formula As a reference benchmark for collaborative learning;

[0251] Reason for prioritizing coefficient λ2 = 0.2:

[0252] Downhole test data shows that this value increases the amount of communication data by ≤15% (compared to the non-cooperative solution), while improving the fault identification rate by 8.7%.

[0253] If λ2 > 0.3, the communication overhead increases sharply by 42% while the accuracy only increases slightly by 1.1%; if λ2 < 0.1, the collaborative learning effect is not significant.

[0254] 2. Step 2: Risk Index Calculation:

[0255]

[0256] Threshold setting:

[0257] R≥8.0: Immediate shutdown (red alert);

[0258] 5.0≤R<8.0: Reduced frequency operation (yellow alert);

[0259] R<5.0: Normal monitoring;

[0260] in:

[0261] R: Real-time dynamic risk index (range 0-10), used to trigger equipment control commands in a graded manner. The calculation frequency is once per second, meeting the response time requirements of coal mine safety regulations;

[0262] P fail The real-time failure probability (0-1) output by the DRL model has the physical meaning of an estimate of the likelihood of equipment failure.

[0263] I severity : Preset fault severity level (1-10), classified according to maintenance cost and safety impact:

[0264] Bearing overheating: I = 3 (repair time ≤ 2 hours);

[0265] Motor insulation failure: I = 7 (downtime loss ≥ ¥200,000);

[0266] Gearbox fracture: I = 10 (potentially causing a gas accumulation accident);

[0267] P power The real-time operating power of the equipment (unit: kW) is derived from the combined calculation of current sensor and voltage transformer.

[0268] 1+log 10 (P power / 1000) represents the principle of logarithmic normalization:

[0269] Eliminate the numerical advantage of high-power equipment (such as a 1500kW coal mining machine) and avoid P fail Minor fluctuations can trigger malfunctions;

[0270] Example: When P power =1500kW, log 10 (1.5)≈0.18, with a denominator of 1.18, significantly weakening the influence of power;

[0271] The constant "1000" is a dimensionless conversion coefficient, making P power / 1000 constitutes a dimensionless ratio;

[0272] Reason: To mitigate the risk of misjudging high-energy-consuming equipment, the threshold was determined based on the ROC curve of 300 failure events.

[0273] Threshold setting basis:

[0274] When R≥8.0, a red alert (immediate shutdown) is triggered: This corresponds to a Class I risk event such as gearbox breakage (based on industry mandatory shutdown standards), with actual accident repair costs exceeding ¥500,000.

[0275] When R < 8.0, a yellow warning (frequency reduction operation) is issued: This covers Class II risks such as bearing overheating. Reducing the frequency to 60% of the rated speed can delay the development of the fault by 2-4 hours, buying time for maintenance.

[0276] When R < 5.0, normal monitoring is performed: only data logs are recorded, and no control actions are triggered to avoid excessive intervention in production.

[0277] It should be noted that industrial tests at a certain coal mine showed that:

[0278] The knowledge-sharing loss function (λ² = 0.2) reduced the major incident underreporting rate to 0% while increasing communication bandwidth utilization by only 14.7%.

[0279] The logarithmic design in the risk index formula reduces the false alarm rate of high-power equipment by 62%, making it more engineering-applicable than the traditional linear formula.

[0280] The graded threshold response mechanism reduced unnecessary shutdowns by 83% and increased coal production by an average of 9,200 tons per year.

[0281] It should be noted that the technical effects of this invention have been verified, and the results are shown in Table 1:

[0282] Table 1 Verification of Technical Effects

[0283] index Traditional methods This invention Increase Fault prediction accuracy 84.2％ 93.6％ An increase of 9.4%. False alarm rate 15.1％ 6.2％ Reduced by 58.9% Average response delay 2.3s 0.8s Reduced by 65.2% Sensor power consumption 100％ 67％ 33% reduction

[0284] Test environment: A coal mine working face, running continuously for 30 days, monitoring 12 pieces of equipment including the coal mining machine and scraper conveyor.

[0285] It is understood that the beneficial effects of the present invention include:

[0286] 1. Closed-loop coupling of DRL and dynamic sampling: through formula θ DRL Humidity attenuation term and f samp The Sigmoid response addresses the problem of precipitous performance degradation in traditional models when downhole humidity changes abruptly.

[0287] 2. Lightweight Blockchain Design: Applying SHA-256 hashing and BFT consensus to resource-constrained mining edge devices to solve the "security-latency" paradox of the Industrial Internet of Things.

[0288] 3. Normalization modeling of the risk index R: Introducing the logarithmic denominator log 10 (P power / 1000) converts kW-level power into a dimensionless risk factor, eliminating the difference in scale between different equipment.

[0289] Please see Figure 2 The present invention provides another embodiment, which provides an intelligent monitoring system for underground equipment in coal mines based on the Internet of Things (IoT). The intelligent monitoring system for underground equipment in coal mines based on the IoT includes:

[0290] The acquisition module 100 includes multiple IoT sensors deployed underground in the coal mine to collect status data of the target device, including vibration data, temperature data and current data.

[0291] The control module 200 includes an edge computing node and a central server installed underground. It is used to preprocess the collected data at the edge computing node, including noise suppression and feature extraction; transmit the preprocessed data to the central server via an Internet of Things (IoT) network; analyze the preprocessed data on the central server using a deep reinforcement learning model to assess equipment status and predict potential faults; generate alarms or control signals based on the prediction results and output them to the execution equipment; and adjust the decision parameters of the deep reinforcement learning model according to real-time environmental feedback to achieve adaptive optimization.

[0292] In a preferred embodiment, this application also provides an electronic device, the electronic device comprising:

[0293] The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, implement the IoT-based intelligent monitoring method for underground equipment in coal mines. The computer device can be broadly categorized as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the method of the present invention.

[0294] This invention can be implemented as a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the methods of embodiments of the invention to be performed. In one embodiment, the computer program is distributed across multiple network-coupled computer devices or processors, such that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be executed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be executed by one or more computer devices or processors, and one or more other method steps / operations may be executed by one or more other computer devices or processors. One or more computer devices or processors may execute a single method step / operation, or execute two or more method steps / operations.

[0295] Those skilled in the art will understand that the method steps of this invention can be performed by a computer program instructing related hardware, such as a computer device or processor, to perform the steps of this invention when executed. Depending on the context, any references herein to memory, storage, databases, or other media may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0296] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0297] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for intelligent monitoring of underground equipment in coal mines based on the Internet of Things, characterized in that, Includes the following steps: S100. Collect status data of the target device through multiple IoT sensors deployed underground in the coal mine. The status data includes vibration data, temperature data, and current data. S200. The edge computing nodes set up downhole preprocess the collected data, and the preprocessing includes noise suppression and feature extraction. The preprocessed data is transmitted to the central server via an Internet of Things (IoT) network. S300: The central server uses a deep reinforcement learning model to analyze preprocessed data to assess equipment status and predict potential failures; based on the prediction results, alarms or control signals are generated and output to the execution device; the decision parameters of the deep reinforcement learning model are adjusted according to real-time environmental feedback to achieve adaptive optimization.

2. The method as described in claim 1, characterized in that, The noise suppression step includes using an adaptive Kalman filter, whose update formula is: in: The state estimate at the current time step represents the filtered sensor data value. The state estimate of the previous time step represents the filtered output of the previous moment; K k Kalman gain matrix, representing adaptively adjusted weights used for prediction and measurement based on noise statistical balance; z k The measured value at the current time step represents the raw sensor data. H: Observation matrix, representing the transformation coefficients from state to measurement.

3. The method as described in claim 1, characterized in that, The feature extraction step includes calculating multimodal time series features, the formula of which is: f(t)=α·FFT(v t )+β·RMS(T t )+γ·I peak (t) in: f(t): The value of the extracted feature vector at time t, representing the feature representation of the fusion of vibration, temperature and current; α, β, γ: Weighting coefficients, representing the relative importance factors of each sensor's data; FFT(v t ): The fast Fourier transform value of vibration data at time t, representing the vibration frequency characteristics; RMS(T t ): The root mean square value of temperature data at time t, representing the characteristics of temperature fluctuation; I peak (t): The peak value of the current data at time t, representing the characteristics of current overload risk.

4. The method as described in claim 1, characterized in that, The deep reinforcement learning model includes an LSTM network and a policy network, wherein the state output formula of the LSTM network is: h t =LSTM(x t ,h t-1 ,c t-1 ) in: h t The hidden state at the current time step represents the time-series characteristics of the device state. x t Input data, representing the preprocessed feature vectors; h t-1 The hidden state of the previous time step represents the memory of the historical state; c t-1 The cell state at the previous time step represents long-term dependency information.

5. The method as described in claim 1, characterized in that, The method also includes using blockchain security mechanisms to verify data transmission, specifically: Generate data hashes for key events at edge nodes; Record the hash value to a lightweight blockchain network; The hash generation formula is: Hash=SHA-256(data+timestamp) in: Hash: Outputs a hash value, representing a unique and immutable identifier for the data; data: Input data, representing compressed sensor data packets; timestamp: A timestamp that represents the absolute time of data collection.

6. The method as described in claim 1, characterized in that, The adaptive optimization step includes dynamically adjusting the sensor sampling frequency based on the environmental humidity parameter, and the adjustment formula is as follows: f samp =f min +(f max -f min )·σ(Δh) in: f samp The adjusted sampling frequency represents the sensor's data acquisition rate. f min Minimum allowed sampling frequency, representing the preset base sampling rate; f max : Maximum allowed sampling frequency, representing the preset maximum sampling rate; σ(Δh): Sigmoid activation function, representing a modulating factor mapped according to the rate of change of humidity Δh (representing the derivative of ambient humidity).

7. The method as described in claim 4, characterized in that, The policy network outputs an action value function, the formula of which is: in: Q(s,a): State-action value, representing the expected long-term reward of taking action a in state s; s: Current state, representing the device state characteristics output by the LSTM; a: Current action, representing a control decision such as "initiating frequency reduction" or "sending an alarm"; r: Instant reward, representing feedback calculated based on prediction accuracy; γ: Discount factor, representing the weight of future returns (0 < γ < 1); s′: Next state, representing the state predicted by the LSTM at the next time step.

8. The method as described in claim 1, characterized in that, The method also includes inter-device collaborative learning, wherein the loss function is: L=λ1L pred +λ2L comm in: L: Total loss value, representing the model optimization objective; L pred : Predicted loss, representing the error in fault prediction; L comm Communication loss represents the overhead of data transmission between devices; λ1, λ2: Weighting parameters, representing the balance between prediction accuracy and communication cost.

9. The method as described in claim 1, characterized in that, The alarm triggering step includes calculating a risk index, the formula of which is: in: R: Risk Index, representing the overall risk score of equipment failure; P fail Fault probability, representing the output value of the deep reinforcement learning model; I severity Severity coefficient: Represents the degree of harm that the fault may cause; E energy Energy consumption value, representing the current power consumption of the device.

10. The method as described in claim 1, characterized in that, The adaptive optimization step includes adjusting the DRL model parameters using an environmental adaptive factor, the formula of which is: i new =θ old ·the·e -α·env_var in: θ new The updated model parameters represent the adjusted weights. θ old Original model parameters, representing the initial weights; η: Learning rate, representing the update step size factor; α: Attenuation coefficient, representing the sensitivity to environmental influences (0 < α < 1); env_var: An environment variable representing the real-time monitored humidity or temperature value.

Images