Multistage safety protection method for weight driving of underground sensor

Abstract

According to the weight-driven multi-stage safety protection method for the underground sensor, the sensor is classified into a core layer, a middle layer and a base layer, the real-time weight is dynamically calculated by fusing the environment anti-interference coefficient, the historical accuracy rate and the conflict punishment, a three-stage verification mechanism is driven, and primary verification is classified according to the weight; the intermediate check takes a high-weight node as a space-time reference, and the LSTM adopts a square weighted loss function; the tampering probability corrected by the random forest output weight is verified in an advanced mode; and a verification result triggers a corresponding tampering response mechanism, and closed-loop protection is realized in combination with federal learning and block chain evidence storage. The problem of data tampering caused by underground electromagnetic interference, equipment aging and hostile attacks can be effectively solved in the whole day, and the reliability and safety of a monitoring system are improved.

Description

Technical Field

[0001] This invention belongs to the field of industrial Internet of Things (IoT) mine safety technology, specifically relating to a multi-level safety protection method driven by the weights of underground sensors. Background Technology

[0002] The working environment in mines is extremely complex, with strong electromagnetic interference, signal attenuation due to physical obstruction, and sensor equipment aging easily due to the harsh environment. These factors not only lead to sensor data distortion but also provide opportunities for malicious attackers to inject forged data. Current security technologies, such as simple threshold alarms, data hash verification, or static redundancy backups, have shortcomings such as high false alarm rates, inability to handle common-cause failures, and neglect of differences in node reliability. They are insufficient to effectively identify highly concealed data tampering while ensuring system real-time performance. Summary of the Invention

[0003] To address the aforementioned problems, the purpose of this invention is to provide a multi-level safety protection method driven by downhole sensor weights, which solves the shortcomings of existing downhole safety monitoring technologies that cannot dynamically assess sensor reliability and are difficult to accurately identify data tampering under complex interference.

[0004] To achieve the above objectives, the technical solution adopted by the present invention includes: A multi-level safety protection method driven by downhole sensor weights includes the following steps: S1 collects sensor data in the core layer, intermediate layer and foundation layer of the coal mine in real time. The sensor data includes at least sensor readings and CRC check codes. Multiple gas sensors are distributed in the core layer, multiple temperature sensors and multiple vibration sensors are distributed in the intermediate layer, and multiple humidity sensors are distributed in the foundation layer. S2 presets the initial trusted dynamic weight for each sensor and performs a preliminary verification on the raw data of all sensor data acquired in S1. If the preliminary verification passes, the data is complete and error-free, marked as "communication complete," and passed to the multi-dimensional feature extraction and intermediate verification stages for analysis. If the preliminary verification fails: bit errors may occur in the data packet during transmission, and the system will trigger a retransmission mechanism with a maximum of 3 retransmissions. If it fails consecutively, the data packet may be discarded, and the trusted dynamic weight of that sensor will be lowered. S3 involves multi-dimensional feature extraction from the sensor data that passed the initial verification in S2, yielding time-domain, frequency-domain, and correlation features. Intermediate verification is then performed on the same sensor data to obtain spatial, temporal, and prediction scores. Finally, LSTM prediction verification is applied to the sensor data that passed the initial verification in S2, adjusting the spatial, temporal, and prediction scores, and calculating the adjusted overall confidence level and sensor conflict penalty term. ; When the adjusted overall confidence level is ≥0.8 and the scores of each item are ≥0.6, the intermediate verification passes and monitoring continues; when the adjusted overall confidence level is ∈[0.6,0.8], or any item score is <0.3, the intermediate verification issues a warning and monitoring is enhanced; when the adjusted overall confidence level is <0.6, the intermediate verification fails and advanced verification is triggered. S4 performs random forest multi-model ensemble inference and confidence calibration and decision-making on the evaluation results of intermediate-level verification failure and triggering advanced-level verification in S3, and obtains the final anomaly probability, decision conclusion and disposal suggestions, and adjusts the credible dynamic weights.

[0005] Preferably, the primary check is CRC32 check.

[0006] Preferably, intermediate-level verification includes spatiotemporal consistency verification.

[0007] Preferably, in S2, multi-dimensional feature extraction is performed on each sensor data, specifically including: S21, for each sensor data collected in S1, calculate the mean μ and variance using a sliding window statistical method. skewness and kurtosis The data is segmented and quantized into 5 levels of symbols. Time-domain features are obtained through approximate entropy calculation, where the time-domain features are the approximate entropy values ​​for each sensor. pEn; S22, each sensor data acquired by S1 is decomposed using wavelet basis functions to obtain the wavelet coefficient energy spectrum. Combined with abnormal frequency bands and sampling frequency The frequency domain characteristics are calculated, and these characteristics represent the proportion of energy in the anomalous frequency band. Abnormal frequency band The characteristic frequency band of known interference or attack is indicated by actual measurement of the electromagnetic spectrum downhole; sampling frequency. Indicates the sensor's nominal sampling rate; S23. Two sets of observation data collected by different types of sensors within the same area and within the same time period are used to obtain temperature measurement sequences and pressure measurement sequences. The Pearson correlation coefficient between the temperature measurement sequences and the pressure measurement sequences is then calculated. .

[0008] Preferably, the approximate entropy value of the sensor When pEn < 0.3, it indicates a low entropy value; the data is too regular and may be manipulated by fixed values ​​or simple patterns, requiring intermediate-level verification. The sensor's approximate entropy value... When pE > 0.8, it is considered a high entropy value, the data is too random and may be subject to noise interference, so intermediate verification is required; when the approximate entropy value of the sensor is ∈ [0.3-0.8], it is within the normal range and reflects the natural fluctuations of the downhole environment.

[0009] Preferably, the proportion of abnormal frequency band energy ∈[0,1], this value will serve as a key feature for subsequent judgment. When it exceeds the preset threshold, it is determined that the current signal is subject to abnormal interference and intermediate verification is prepared.

[0010] Preferably, when the Pearson correlation coefficient is... When the Pearson correlation coefficient is 1, it indicates a perfect positive correlation between temperature and pressure changes; when the Pearson correlation coefficient is 1, it indicates a perfect positive correlation between temperature and pressure changes. A Pearson correlation coefficient ≥ 0.7 indicates a perfect positive correlation between temperature and pressure changes; when the Pearson correlation coefficient is ≥ 0.7, it indicates a perfect positive correlation between temperature and pressure changes. If the value is less than 0.7, it is determined that the sensor has encountered an abnormality or other extreme or physically inconsistent special operating conditions, and intermediate calibration is to be performed.

[0011] Preferably, the adjusted overall confidence level in S3 is calculated as follows: 0.4 × spatial score adjustment value + 0.3 × time score adjustment value + 0.3 × prediction score adjustment value.

[0012] Preferably, the reliable dynamic weight of each sensor is calculated using equation (1). ; (1) in, This represents the baseline weight of the sensors at each level; the baseline weight of the gas sensors in the core layer is... The value ranges from 0.8 to 1.0; the reference weights for the temperature and vibration sensors in the intermediate layer are... The value ranges from 0.5 to 0.7; the baseline weight of the humidity sensor in the base layer is... The value ranges from 0.1 to 0.4; among which, Indicates the environmental interference immunity coefficient. SNR is the signal-to-noise ratio; where, Indicates historical accuracy. ;in, This represents a data conflict penalty, which adds +0.1 per instance when the difference between sensors of the same type exceeds a threshold. The value ranges from 0.0 (no conflict) to 0.5 (severe conflict).

[0013] Preferably, the trusted dynamic weights of each sensor The range of values ​​is .

[0014] Compared with the prior art, the advantages of the present invention are: (1) A multi-level safety protection method driven by downhole sensor weights, which integrates the physical layer environmental disturbance immunity coefficient, the data layer historical accuracy and conflict penalty through a dynamic weight model to achieve real-time quantitative evaluation of sensor reliability. Combined with a three-level verification mechanism, especially the weighted LSTM prediction and spatiotemporal consistency analysis in the intermediate verification, it can effectively identify abnormal scenarios that are difficult to detect by traditional methods, such as periodic masking attacks, low-amplitude drift and common-cause faults, and significantly reduce the false alarm rate and false negative rate.

[0015] (2) A multi-level safety protection method driven by downhole sensor weights introduces a "core layer weight priority" mechanism, which prioritizes high-weight nodes during data transmission and processing, reduces redundant calculation and communication overhead, and improves transmission efficiency and network fault tolerance while ensuring data reliability. The overall resource consumption of the system is significantly reduced compared with traditional solutions.

[0016] (3) A multi-level security protection method driven by downhole sensor weights. For extreme environments such as high temperature, high humidity, and strong electromagnetic interference in downhole systems, an environment-adaptive dynamic baseline correction and frequency domain feature analysis mechanism is designed. This mechanism can adjust the detection threshold and feature weights according to real-time operating conditions, avoiding false alarms caused by environmental changes and improving the stability and reliability of the system in practical applications. It possesses a complete tamper response and weight update mechanism, which can adjust sensor weights in real time based on detection results and implement measures such as isolation, data interpolation, or activation of backup equipment for confirmed abnormal nodes. Through continuous optimization and hot model updates via federated learning, the system has self-evolution capabilities and can quickly adapt to new attack patterns and environmental changes. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the following detailed description to explain the invention, but do not constitute a limitation thereof. In the drawings: Figure 1 Schematic diagram of real-time response measures. Detailed Implementation

[0018] The invention is not limited to the specific embodiments described below. All equivalent modifications made based on the technical solutions of this application fall within the protection scope of this invention. Unless otherwise specified, all components and devices in this invention utilize components and devices known in the prior art.

[0019] Example 1 This embodiment discloses a multi-level safety protection method driven by downhole sensor weights, which specifically includes the following steps: S1, Real-time acquisition of sensor data in the core layer, intermediate layer and foundation layer of underground coal mine, wherein the sensor data includes at least sensor readings and CRC check codes; The core layer contains multiple gas sensors, the middle layer contains multiple temperature sensors and multiple vibration sensors, and the base layer contains multiple humidity sensors. Multi-source data acquisition is fundamental: it is responsible for acquiring the most raw data from various downhole sensors and assigning an initial dynamic weight based on the sensor type and initial state. This is the data source and starting point of the entire system.

[0020]

[0021] S2: Preset the initial reliable dynamic weight for each sensor and perform a primary verification on the raw data of all sensor data acquired in S1; in this embodiment, the primary verification is CRC32 verification. In fact, the primary verification in this embodiment includes CRC and multi-dimensional feature extraction. First, the data is CRC verified; if the verification fails, the data is discarded. Then, multi-dimensional feature extraction is performed on the CRC-verified data. If an anomaly is detected by the multi-dimensional feature extraction, the information is passed to the intermediate verification as a feature of the intermediate verification. It is important to clarify that multi-dimensional feature anomaly detection is only a step in the primary verification; it does not directly trigger the intermediate verification. Instead, all data must undergo intermediate verification, and the result of multi-dimensional feature anomaly detection is an important input to the intermediate verification (the intermediate verification can only be started after all inputs are available; see the section on intermediate verification inputs below for details).

[0022] If the initial verification passes, the data is complete and error-free, marked as "communication complete," and passed to the multi-dimensional feature extraction and intermediate verification stages for analysis. If the initial verification fails, bit errors may occur in the data packet during transmission, and the system will trigger a retransmission mechanism (maximum number of retransmissions is 3). If it fails consecutively, the data packet may be discarded, and the reliable dynamic weight of the sensor will be lowered. In S2 of this embodiment, multi-dimensional feature extraction is performed on each sensor data, specifically including: S21, for each sensor data collected in S1, calculate the mean μ and variance using a sliding window statistical method. skewness and kurtosis The data is segmented and quantized into 5 levels of symbols, and the temporal features are obtained through approximate entropy calculation. The temporal features are the approximate entropy values ​​of each sensor. pEn; The approximate entropy value of the sensor When pEn < 0.3, it is considered a low entropy value, and the data is too regular, potentially manipulated by fixed values ​​or simple patterns. Intermediate verification is required. The approximate entropy value of the sensor... When pE > 0.8, it is considered a high entropy value, the data is too random and may be subject to noise interference, so intermediate verification is required; when the approximate entropy value of the sensor is ∈ [0.3-0.8], it is within the normal range and reflects the natural fluctuations of the downhole environment.

[0023] The approximate entropy calculation in this embodiment was proposed by Steven M. Pincus in 1991 to quantify the complexity and regularity of time series; Pincus, SM (1991). Approximate entropy as a measure of system complexity. Proceedings of the National Academy of Sciences, 88(6), 2297–2301.

[0024] S22, each sensor data acquired by S1 is decomposed using wavelet basis functions to obtain the wavelet coefficient energy spectrum. Combined with abnormal frequency bands and sampling frequency The frequency domain characteristics are calculated, and these characteristics represent the proportion of energy in abnormal frequency bands. The abnormal frequency band The sampling frequency represents the characteristic frequency band of known interference or attack, measured from the downhole electromagnetic spectrum; Indicates the sensor's nominal sampling rate; percentage of energy in abnormal frequency bands. ∈[0,1], this value will serve as a key feature for subsequent judgment. When it exceeds the preset threshold, it is determined that the current signal is subject to abnormal interference and intermediate verification is prepared.

[0025] In this embodiment, at the coal mining face, the frequency converter of the coal mining machine and the high-power scraper conveyor will continuously emit strong harmonic interference of 50-100Hz, which is like launching electronic smoke bombs in the monitoring frequency band. Traditional time-domain detection (such as mean abrupt changes) has a high failure rate in this environment because it cannot distinguish between real gas outbursts and motor interference pulses. This invention specifically extracts the energy proportion characteristics of abnormal frequency bands and accurately locates the interference frequency band information unique to coal mine electromechanical equipment. When high-frequency cutting operations generate abnormal harmonics, the system can immediately identify the anomaly in the background noise.

[0026] Among them, the proportion of abnormal frequency band energy This is a normalized ratio value between 0 and 1. It represents the proportion of energy within the abnormal frequency band to the total signal energy. The closer the value is to 1, the more concentrated the signal energy is in the abnormal interference frequency band, and the greater the possibility of the signal being interfered with by human or environmental factors; the closer the value is to 0, the more the signal energy is mainly distributed in the normal frequency band, and the higher the signal reliability. (The sensor signal is a low-frequency signal and may overlap with interference signals in the 50-100Hz range). Wavelet coefficient energy spectrum This represents the wavelet coefficients at various scales (corresponding to different frequency bands) obtained after multi-resolution decomposition of the original time-domain vibration signal using the db4 wavelet basis function. The square of the modulus is the energy spectrum. This is an array describing the distribution of signal energy at different frequency points and is the foundation of the entire calculation. Abnormal frequency bands. This is an engineering setting based on prior knowledge. In this application scenario, it specifically refers to the strong harmonic interference frequency band generated by high-power equipment such as coal mining machines and frequency converters, such as motor harmonics in the 50-100Hz range. Sampling frequency These are the hardware parameters of the data acquisition system.

[0027] S23. Two sets of observation data collected by different types of sensors within the same area and within the same time period are used to obtain temperature measurement sequences and pressure measurement sequences. The Pearson correlation coefficient between the temperature measurement sequences and the pressure measurement sequences is then calculated. .

[0028] In this embodiment, when the Pearson correlation coefficient is... When the Pearson correlation coefficient is 1, it indicates that the changes in temperature and pressure are completely positively correlated; when the Pearson correlation coefficient is 1, it indicates that the changes in temperature and pressure are completely positively correlated. A Pearson correlation coefficient ≥ 0.7 indicates a perfect positive correlation between temperature and pressure changes; when the Pearson correlation coefficient is... If the value is less than 0.7, it is determined that the sensor has encountered an abnormality or other extreme or physically inconsistent special operating conditions, and intermediate calibration is to be performed.

[0029] S3 performs multi-dimensional feature extraction on the sensor data that passed the initial verification in S2, obtaining time-domain features, frequency-domain features, and correlation features. These time-domain features, frequency-domain features, and correlation features are then "translated" into a series of quantifiable feature indicators representing different dimensions (such as mean μ, variance σ) using commonly used methods. 2 (Approximate entropy ApEn, abnormal frequency band energy ratio E_abn, correlation coefficient r_Tp, etc.); intermediate verification is performed on the sensor data that passed the primary verification in S2 to obtain spatial score, temporal score and prediction score; the intermediate verification in this embodiment includes spatiotemporal consistency verification.

[0030] LSTM prediction verification is performed on the sensor data that passed the initial verification in S2, adjusting the spatial score, temporal score, and prediction score. The adjusted overall confidence level and sensor conflict penalty term are then calculated. The adjusted overall confidence level = 0.4 × spatial score adjustment value + 0.3 × time score adjustment value + 0.3 × prediction score adjustment value.

[0031] S3.1 First Phase: Data Preparation and Benchmark Calculation Processing objective: To establish spatial and temporal benchmarks. Specific steps: 1. Spatial reference calculation The "normal value range" of neighboring nodes is calculated using the weighted average method. Formula: Spatial baseline value = Σ (neighbor node values ​​× normalized weights); Normalized weight = Weight of a single neighboring node / Sum of weights of all neighboring nodes Logic: Readings from high-weight nodes have a greater impact on the benchmark value. 2. Time base calculation Statistical characteristics were calculated based on the previous 59 historical values ​​from the current sensor. Includes: historical mean, historical standard deviation, and data trend. Logic: Establish the normal behavior pattern of this sensor. S3.2 Second Phase: Spatial Consistency Verification Processing objective: To verify whether the current readings conform to a spatial distribution pattern. Verification process: 1. Deviation Calculation Calculate the absolute difference between the current reading and the spatial reference value. Formula: Spatial Deviation = |Current Reading - Spatial Reference Value| Normalized spatial bias = |current reading - weighted average of neighboring nodes| / standard deviation of neighboring node readings 2. Threshold Comparison Use a fixed spatial threshold (e.g., 0.1) that is not adjusted according to environmental changes. Comparison: Spatial Bias and Fixed Threshold 3. Spatial score calculation If the deviation is less than or equal to the threshold: Spatial score = 1.0 If deviation > threshold: Spatial score = threshold / deviation Logic: The larger the deviation, the lower the score, with proportional decay. S3.3 Phase 3: Time Consistency Verification Processing objective: To verify whether the current readings conform to the time series pattern. Verification process: 1. Statistical anomaly detection Calculate the Z-score (standard score) of the current reading compared to the historical mean. Formula: Z = |Current value - Historical mean| / Historical standard deviation 2. Trend Continuity Analysis Check if the current reading is consistent with the direction of the data trend. Avoid sudden reversals or abnormal amplitudes 3. Time Score Calculation Calculated based on Z-score: Time score = max(0, 1-Z / 5) Logical reasoning: The higher the Z-score, the greater the degree of anomaly, and the lower the score.

[0032] The score comparison table is as follows:

[0033] S3.4 Fourth Stage: LSTM Prediction Verification Processing objective: To validate the reasonableness of the data using machine learning models. Verification process: 1. Sequence prediction Using an LSTM model, predict the 61st value based on the first 60 values. The model has learned the normal variation patterns of the sensor. 2. Calculation of prediction error Formula: Prediction Error = |Actual Value - Predicted Value| 3. Prediction Score Calculation Based on the comparison of error distribution with historical error distribution If the prediction error is ≤3 times the historical error standard deviation: score = 1.0 If the prediction error is greater than 3 times the standard deviation: Score = (3 × standard deviation) / error S3.5 Comprehensive Decision Making When the adjusted overall confidence level is ≥0.8 and the scores of each item are ≥0.6, the intermediate verification passes and monitoring continues; when the adjusted overall confidence level is ∈[0.6,0.8], or any item score is <0.3, the intermediate verification issues a warning and monitoring is enhanced; when the adjusted overall confidence level is <0.6, the intermediate verification fails and advanced verification is triggered. S4 performs random forest multi-model ensemble inference and confidence calibration and decision-making on the evaluation results of intermediate-level verification failure and triggering advanced-level verification in S3, and obtains the final anomaly probability, decision conclusion and disposal suggestions, and adjusts the credible dynamic weights.

[0034] S4.1 First Stage: Feature Fusion Processing objective: To integrate multi-source features into an input vector that the model can understand, and then feed it into the random forest model. Specific steps: 1. Feature standardization Normalize all values ​​to the same scale (0-1 range) to prevent certain features from dominating model decisions due to excessively large values. Formula: Standardized value = (Original value - Sample mean) / Sample standard deviation Result: The data is transformed into a distribution with a mean of 0 and a standard deviation of 1. 2. Cross-feature generation Create interaction terms between features to discover non-linear relationships between features. Spatial anomaly weight adjustment: Spatial score × sensor weight, spatial anomalies of nodes with high weights are more important. Anomaly Significance Ratio: Spatial Bias / Temporal Variance, the degree of significance of anomalies relative to normal fluctuations. Spatiotemporal consistency discrepancy: |spatial score - temporal score|, consistency of different testing methods 3. Temporal Feature Extension Trend characteristics generated from historical data: capturing the time evolution pattern of data such as recent rate of change and volatility. Short-term trend: The direction of data change in the last hour (upward / downward / stable) For example, the gas concentration rose from 0.5% to 1.8% in the last hour, showing a strong upward trend. Rate of change: How quickly a value changes For example: an increase of 0.02% per minute shows a change of 0.02. Periodic comparison: Differences from data at the same time yesterday For example: Yesterday at this time, it was 0.5%, and today it is 1.8%, showing a difference of 1.3. Duration of abnormality: How long the abnormal state has lasted. For example: the abnormal time lasted for 45 minutes. 4. Input vectors that the model can understand

[0035] S4.2 Random Forest Multi-Model Ensemble Inference Processing objective: To obtain robust anomaly probability assessments through ensemble learning. Reasoning process: 1. Multi-tree parallel prediction Each decision tree in a random forest independently classifies the feature vectors. Each tree is judged based on a randomly selected subset of features. 2. Voting Mechanism Each tree outputs an "anomaly probability" (a value between 0 and 1). Final probability = average of the probabilities output by all trees 3. Probability Adjustment The anomaly detection of high-weighted sensors is more stringent. Formula: Adjusted probability = Original probability × (1 + Weighting penalty factor) Weight penalty factor = (Current sensor weight - Trusted dynamic weight) × Adjustment intensity coefficient Current sensor weight: 0.1-1.0, reflecting the sensor's historical reliability. Trustworthy dynamic weight: Usually set to 0.5 (medium reliability level) Adjust the intensity coefficient: This controls the adjustment range, typically set to 0.4-0.6. S4.3 Phase 3: Confidence Calibration and Decision Making Processing objective: To transform the probabilities output by the model into actionable decisions. Calibration process: 1. Probability distribution mapping Based on historical validation data, a correlation between probability and true anomaly rate is established to correct systematic biases in the model. a. Collect historical data The model's past predictions of 1000 anomaly probabilities Are the actual results corresponding to these predictions truly abnormal?

[0036] b. Establish corresponding relationships c. Create correction rules For example: if the model predicts a probability between 80% and 90%, change it to 70%. If the model predicts a probability between 90% and 100%, adjust it to 85%. ... ... 2. Multi-threshold decision Set three-level decision boundaries: Safe zone (probability < 0.3): Data is reliable. Warning zone (0.3 ≤ probability < 0.7): Manually verify or mark as "pending confirmation" or observe at several more time points. Danger zone (probability ≥ 0.7): Confirmed as abnormal. 3. Quantification of Uncertainty Calculate the degree of uncertainty in the model's predictions; when uncertainty is high, conservative decision-making is recommended.

[0037] Voting mechanism of random forest: Suppose 100 trees independently evaluate the same data; Each tree was voted "abnormal" or "normal"; Final probability = the proportion of trees that are "abnormal".

[0038] When the proportion of trees with "abnormal" ratings is greater than 70% out of 100 trees, it is a deterministic state, and the result can be trusted, thus confirming the anomalous nature of the tree and placing it in a danger zone.

[0039] When the percentage of trees with "abnormal" results out of 100 trees is less than 30%, it is a deterministic state, and the result can be trusted, confirmed as normal, and is in a safe zone.

[0040] When the proportion of trees with "abnormal" values ​​among 100 trees is between 30% and 70%, it is a state of high uncertainty, requiring a conservative approach and falling within the warning zone. The handling strategy at this time is as follows: Add a review step: do not process immediately, mark as "pending confirmation" first; Extend the observation period: Observe the data at multiple time points before making a decision; Manual review.

[0041] In this embodiment, the reliable dynamic weight of each sensor is calculated using equation (1). ; (1) in, This represents the baseline weight of the sensors at each level; the baseline weight of the gas sensors in the core layer is... The value ranges from 0.8 to 1.0; the reference weights for the temperature and vibration sensors in the intermediate layer are... The value ranges from 0.5 to 0.7; the baseline weight of the humidity sensor in the base layer is... The value ranges from 0.1 to 0.4; in, Indicates the environmental interference immunity coefficient. SNR is the signal-to-noise ratio; in, Indicates historical accuracy. ; in, This represents a data conflict penalty, which is +0.1 per instance when the difference between sensors of the same type exceeds a threshold. The value ranges from 0.0 (no conflict) to 0.5 (severe conflict).

[0042] The reliable dynamic weight of each sensor in this embodiment The range of values ​​is .

[0043] Example 2 This embodiment discloses a case study of monitoring an abnormal surge in methane concentration in a coal mine return airway. 1. Scene setting (1) Location: Return airway of a coal mine (2) Event: Injection of fake data (3) Initial environment (09:30:00)

[0044] Environmental conditions: Motor startup caused SNR=14dB (strong interference).

[0045] Equipment status: The coal mining machine is operating normally (vibration threshold > 0.3g during normal operation). Permissible gas fluctuation: ≤0.5% / min during normal operation of the coal mining machine. a. Dynamic weight calculation (before an anomaly) S001: →1.0 (Truncation) S002: →1.0 (Truncation) S003: →1.0 (Truncation) S004:

[0046] S005: →1.0 (Truncation) b. Abnormal events Abnormal action: Data in S001 and S003 has been tampered with. Real-time reading (09:30:15):

[0047] c. Anomaly Detection Process ① Initial verification (09:30:16) Verify data integrity (complete within 5ms)

[0048] ② Anomaly detection based on independent analysis of historical data (09:30:18)

[0049] System flags: SOO1 Time series anomaly (high risk), S003 Threshold exceedance anomaly (high risk) ③ Spatial consistency check (09:30:21) 1) Methane sensor cross-validation Reference node selection: S001 ( S002 ) Deviation calculation:

[0050] Allowable threshold:

[0051] Deviation (2.85) > Allowable threshold (0.1) → At least one of S001 and S002 is distorted.

[0052] 2) Vibration sensor cross-validation Reference node selection: S003 ( S005 ) Deviation calculation:

[0053] Allowable threshold:

[0054] Deviation (2.25) > Allowable threshold (0.1) → At least one of S003 and S005 is distorted.

[0055] ④ Verification of physical laws (09:30:25) Verify the physical compatibility between equipment status and gas changes.

[0056] 1) Supplementary verification: S005 vibration value 0.65g > 0.3g → The actual equipment is still running.

[0057] The SO2 gas value of 0.41% is lower than the previous value of 0.57%, which is consistent with the trend of normal dilution of gas concentration in underground ventilation systems.

[0058] 2) Final judgment: S003 contradicts the physical reality that the device is in operation.

[0059] The dramatic increase in gas in S001 contradicts the dilution trend in S002.

[0060] ⑤ Decision-making level integration judgment (09:30:28) 1) Input feature matrix

[0061] 2) Random Forest Output

[0062] d. Response and Weight Update ① Real-time response measures such as Figure 1 As shown.

[0063] ② Dynamic weight update 1) Update rules:

[0064]

[0065] 2) Punishment / Reward Mechanism: Anomaly confirmed: β_conf increased (0.3 / 0.2), α_his decreased.

[0066] Validation contribution: α_his increased (0.05 / 0.08).

[0067] Weight truncation: upper limit 1.0, lower limit 0.1.

[0068] The preferred embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. However, this disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this disclosure, various simple modifications can be made to the technical solutions of this disclosure, and these simple modifications all fall within the protection scope of this disclosure.

[0069] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0070] Furthermore, the various implementation methods disclosed in this solution can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content invented by this disclosure.

Claims

1. A multi-level safety protection method driven by downhole sensor weights, characterized in that, Includes the following steps: S1, Real-time acquisition of sensor data in the core layer, intermediate layer and foundation layer of underground coal mine, wherein the sensor data includes at least sensor readings and CRC check codes; The core layer contains multiple gas sensors, the middle layer contains multiple temperature sensors and multiple vibration sensors, and the base layer contains multiple humidity sensors. S2, preset the initial reliable dynamic weight of each sensor, and perform a preliminary verification of the raw data of all sensor data obtained in S1; If the initial verification passes, the data is complete and error-free, marked as "communication complete", and passed to the multi-dimensional feature extraction and intermediate verification stages for analysis; If the initial verification fails: the data packet may have bit errors during transmission, and the system will trigger a retransmission mechanism with a maximum of 3 retransmissions; if it fails consecutively, the data packet may be discarded and the trusted dynamic weight of the sensor will be reduced. S3, multi-dimensional feature extraction is performed on the sensor data that passed the initial verification in S2 to obtain time-domain features, frequency-domain features and correlation features; The sensor data that passed the initial verification in S2 are subjected to intermediate verification to obtain spatial score, temporal score and prediction score; LSTM prediction verification is performed on the sensor data that passed the initial verification in S2, adjusting the spatial score, temporal score, and prediction score. The adjusted overall confidence level and sensor conflict penalty term are then calculated. ; When the adjusted overall confidence level is ≥0.8 and the scores of each sub-item are ≥0.6, the intermediate verification is passed and monitoring continues. If the adjusted overall confidence level is ∈ [0.6, 0.8], or any score is < 0.3, a mid-level verification warning is issued, and enhanced monitoring is implemented. When the adjusted overall confidence level is less than 0.6, the intermediate check fails and the advanced check is triggered. S4 performs random forest multi-model ensemble inference and confidence calibration and decision-making on the evaluation results of intermediate-level verification failure and triggering advanced-level verification in S3, and obtains the final anomaly probability, decision conclusion and disposal suggestions, and adjusts the credible dynamic weights.

2. The multi-level safety protection method driven by downhole sensor weights as described in claim 1, characterized in that, The primary verification is CRC32 verification.

3. The multi-level safety protection method driven by downhole sensor weights as described in claim 1, characterized in that, The intermediate-level verification includes spatiotemporal consistency verification.

4. The multi-level safety protection method driven by downhole sensor weights as described in claim 1, characterized in that, In step S2, multi-dimensional feature extraction is performed on each sensor data, specifically including: S21, for each sensor data collected in S1, calculate the mean μ and variance using a sliding window statistical method. skewness and kurtosis The data is segmented and quantized into 5 levels of symbols, and the temporal features are obtained through approximate entropy calculation. The temporal features are the approximate entropy values ​​of each sensor. pEn; S22, each sensor data acquired by S1 is decomposed using wavelet basis functions to obtain the wavelet coefficient energy spectrum. Combined with abnormal frequency bands and sampling frequency The frequency domain characteristics are calculated, and these characteristics represent the proportion of energy in abnormal frequency bands. ; The abnormal frequency band The characteristic frequency bands of known interference or attacks are represented by actual measurements of the electromagnetic spectrum in the well. The sampling frequency Indicates the sensor's nominal sampling rate; S23. Two sets of observation data collected by different types of sensors within the same area and within the same time period are used to obtain temperature measurement sequences and pressure measurement sequences. The Pearson correlation coefficient between the temperature measurement sequences and the pressure measurement sequences is then calculated. .

5. The multi-level safety protection method driven by downhole sensor weights as described in claim 4, characterized in that, The approximate entropy value of the sensor When pEn < 0.3, it is considered a low-entropy value, and the data is too regular, which may be tampered with by fixed values ​​or simple patterns. It is time to perform intermediate verification. The approximate entropy value of the sensor When pE > 0.8, it is considered a high entropy value, the data is too random and may be subject to noise interference, so intermediate verification is required. When the approximate entropy value of the sensor is ∈ [0.3-0.8], it is within the normal range and reflects the natural fluctuations of the downhole environment.

6. The multi-level safety protection method driven by downhole sensor weights as described in claim 4, characterized in that, The percentage of energy in the abnormal frequency band ∈[0,1], this value will serve as a key feature for subsequent judgment. When it exceeds the preset threshold, it is determined that the current signal is subject to abnormal interference and intermediate verification is prepared.

7. The multi-level safety protection method driven by downhole sensor weights as described in claim 4, characterized in that, When the Pearson correlation coefficient When =1, it indicates that the changes in temperature and pressure are completely positively correlated; When the Pearson correlation coefficient A value ≥0.7 indicates a perfect positive correlation between temperature and pressure changes; When the Pearson correlation coefficient If the value is less than 0.7, it is determined that the sensor has encountered an abnormality or other extreme or physically inconsistent special operating conditions, and intermediate calibration is to be performed.

8. The multi-level safety protection method driven by downhole sensor weights as described in claim 1, characterized in that, The adjusted overall confidence level in S3 is calculated as follows: 0.4 × spatial score adjustment value + 0.3 × time score adjustment value + 0.3 × prediction score adjustment value.

9. The multi-level safety protection method driven by downhole sensor weights as described in any one of claims 1-8, characterized in that, The reliable dynamic weight of each sensor is calculated using equation (1). ; （1） in, This represents the baseline weight of the sensors at each level; the baseline weight of the gas sensors in the core layer is... The value ranges from 0.8 to 1.0; the reference weights for the temperature and vibration sensors in the intermediate layer are... The value ranges from 0.5 to 0.7; the baseline weight of the humidity sensor in the base layer is... The value ranges from 0.1 to 0.4; in, Indicates the environmental interference immunity coefficient. SNR is the signal-to-noise ratio; in, Indicates historical accuracy. ; in, This represents a data conflict penalty, which adds +0.1 per instance when the difference between sensors of the same type exceeds a threshold. The value ranges from 0.0 (no conflict) to 0.5 (severe conflict).

10. The multi-level safety protection method driven by downhole sensor weights as described in claim 9, characterized in that, The reliable dynamic weights of each sensor The range of values ​​is .

Images