Rock mass information fault-tolerant identification method and system based on dynamic weighting of multi-source while-drilling parameter credibility

Abstract

This invention discloses a fault-tolerant identification method and system for rock mass information based on dynamic weighting of multi-source drilling parameter reliability. The method includes: acquiring thrust, torque, rotational speed, and drilling displacement / speed to identify the current process; calculating a four-dimensional score for the integrity, signal stability, physical consistency, and historical health of data from each sensor channel, calculating dynamic reliability and smoothing it; identifying abnormal states based on the smoothed reliability and performing fault-tolerant processing; extracting features from each channel, calculating dynamic fusion weights based on smoothed reliability, channel information gain, and channel redundancy, and constructing fusion features; inputting the fusion features into a rock mass information identification model to output identification results, calculating the reliability of the results, and performing normal output, low-reliability review prompts, or stopping deterministic identification based on the reliability. This invention maintains system availability even under conditions of partial sensor anomalies or failures, reducing the risk of misjudgment caused by erroneous data.

Description

Technical Field

[0001] This invention relates to the fields of intelligent drilling in geotechnical engineering, multi-source information fusion while drilling, and sensor fault diagnosis, and particularly to a fault-tolerant identification method for rock mass information. Background Technology

[0002] Drilling-while-drilling (DWD) rock mass information identification continuously acquires information on the interaction between the drilling rig and the rock mass through sensors such as thrust, torque, rotational speed, drilling displacement, vibration, and motor current. It then uses empirical or intelligent models to identify lithology, rock interface, and rock mechanical parameters. Multi-source DWD information can reflect rock mass characteristics from different perspectives, including load, motion, energy, and vibration, and is a crucial foundation for improving the accuracy and applicability of on-site identification.

[0003] The drilling environment is typically characterized by water mist, dust, impact, vibration, electromagnetic interference, oil temperature variations, and long-distance cabling. Sensors are prone to increased noise, zero-point drift, isolated spikes, short-term disconnections, continuous data loss, range saturation, and communication packet loss. Existing drilling identification systems typically process data before it enters the model using uniform filtering or simple outlier removal, without adjusting the data usage based on the real-time health status of each sensor and the current drilling process. When a key parameter in a particular sensor becomes abnormal, the identification model may still accept erroneous data as normal input, leading to high-confidence errors in lithology, interface, or strength predictions.

[0004] Some multimodal models use fixed channel weights during training or automatically assign feature weights using attention mechanisms. However, the attention weights within the model reflect the degree of contribution to the current recognition task and are not equivalent to the measurement reliability of the sensors. They cannot directly indicate whether the sensors are drifting, disconnected, or stuck, nor can they explicitly select repair, reconstruction, or degraded operation strategies. Therefore, a method is needed that can explicitly evaluate the reliability of each drilling signal and dynamically adjust the fusion weights while maintaining the availability of the recognition system under conditions of partial sensor anomalies or failures.

[0005] Anomalies in multi-source drilling signals exhibit a clear process-dependent nature. For example, a reverse change in displacement during drill bit lifting is normal, but a sudden reversal during stable drilling may indicate displacement sensor jitter or power head rebound; a rapid increase in torque when the drill bit contacts the bottom of the hole is a normal contact response, but a sustained increase during idling may indicate zero-point drift or a transmission system malfunction. Using fixed thresholds that do not differentiate between processes can easily misjudge normal process transitions as sensor failures, or lead to missed detections when a genuine fault occurs due to overly broad thresholds.

[0006] Existing simple filtering methods can typically only handle isolated pulses and cannot handle slow drift, stuck values, and range saturation. Simply using neural networks to reconstruct missing signals may produce superficially smooth but physically incorrect reconstructed values ​​when the current operating conditions exceed the training range. Especially when critical channels such as torque and propulsion fail, if the system does not simultaneously provide reconstruction uncertainty and applicability, subsequent lithology identification models may generate high confidence in erroneous reconstructed values, making it difficult to meet the requirements of traceability and security in engineering sites.

[0007] Furthermore, multiple signals contain both complementary and redundant information. Thrust, torque, drilling speed, and motor current are highly correlated under certain operating conditions; simply normalizing weights based on confidence level could amplify the same error across multiple related channels. While high-frequency signals such as vibration and acoustics have significant noise, they possess independent value in identifying pre-existing fractures, cavities, or stuck drill bits. Therefore, beyond measuring confidence level, it is necessary to consider channel information gain, channel redundancy, and the importance of the current identification task, and to implement a fault-tolerant identification process through explicit physical consistency residuals and uncertainty gating. Summary of the Invention

[0008] In view of this, the present invention provides a fault-tolerant identification method and system for rock mass information based on dynamic weighting of multi-source drilling parameter confidence, in order to solve the technical problems in existing drilling identification technology, such as the inability to distinguish between unreliable sensor measurements and low feature contributions, the inability to perform differentiated fault-tolerant processing for process-dependent anomaly types, the lack of uncertainty assessment in cross-sensor reconstruction leading to the adoption of erroneous reconstruction values ​​with high confidence, and the failure to consider channel information gain and redundancy during multi-source information fusion, which leads to error amplification.

[0009] This invention relates to a fault-tolerant identification method for rock mass information based on dynamic weighting of multi-source drilling parameter reliability, comprising the following steps:

[0010] 1) Synchronously collect thrust, torque, rotational speed, drilling displacement, and drilling speed during the drilling process, and form the k-th multi-source sliding window X according to a unified time reference. k It identifies the current drilling operation based on the thrust, torque, rotational speed, and the direction and duration of drilling displacement or drilling speed. k ;

[0011] 2) For the i-th sensor, calculate the data integrity score respectively. Signal stability score Multi-parameter physical consistency scoring and historical health score Among them, the multi-parameter physical consistency score According to the current drilling procedure s kCalculate the physical relationship residuals between the following drilling parameters; calculate the dynamic reliability of the i-th sensor according to the following formula. :

[0012]

[0013] in, For a function that maps the input to the interval between 0 and 1, a, b, c, d, and e are preset parameters or parameters determined by calibration data;

[0014] The dynamic confidence level is smoothed over time to obtain the smoothed confidence level. ;

[0015] 3) Identify abnormal states based on the magnitude of the smoothing confidence level and its changing trend and rate of change between adjacent sliding windows. Confirm isolated jumps, short-term missing values, persistent drift, stuck values, saturation, and complete failure states based on the duration of the anomaly. Perform fault-tolerant processing on these abnormal states, including: outlier repair for isolated jumps, cross-sensor reconstruction for short-term missing values, baseline correction for persistent drift, reduction of dynamic confidence level or channel masking for stuck values ​​and saturation, and channel masking for complete failure. The cross-sensor reconstruction outputs at least the estimated abnormal channel value, reconstruction uncertainty, and reconstruction applicability.

[0016] 4) Extract the features of each channel after fault tolerance processing. Based on smoothness confidence Channel information gain and channel redundancy The final dynamic fusion weight is calculated using the following formula. :

[0017]

[0018] in, and These are the channel information gain and channel redundancy of the j-th channel, respectively. Number of sensor channels , and This is the adjustment coefficient;

[0019] Constructing fusion features ;

[0020] 5) The fusion features Input a pre-trained rock mass information recognition model, which is used to output at least one of the following recognition results based on fused features: lithology category, rock layer interface location, and rock mechanical parameters, as well as prediction entropy, and calculate the reliability of the result according to the following formula. :

[0021]

[0022] in, The degree of prediction certainty is determined based on the prediction entropy. Weighted sum of the credibility of effective channels, To ensure overall physical consistency of the data, To reconstruct uncertainty, For sample distribution shift, For effective channel redundancy; to The weight is ; clip indicates that the result is restricted to the range of 0 to 1; based on the confidence level of the result, normal output, low confidence review prompt, or stop deterministic identification is performed.

[0023] Furthermore, the data integrity score Based on the missing rate of the i-th sensor within the current sliding window Packet loss rate Range out-of-bounds rate saturation rate and continuous card value ratio Calculate using the following formula:

[0024]

[0025] in, to Non-negative weights;

[0026] The signal stability score Based on the local noise ratio of the i-th sensor within the current sliding window , jump rate Drift slope Periodic interference energy and the proportion of abnormal peak values Calculate using the following formula:

[0027]

[0028] in, to Non-negative weights;

[0029] The multi-parameter physical consistency score The residuals between the physical relationships of the drilling parameters corresponding to the i-th sensor are calculated using the following formula:

[0030]

[0031] in, For the current drilling process, including the first The set of physical relationships of road sensors. For the corresponding physical relationship residuals, For health standards, The weights are non-negative.

[0032] Furthermore, the set of physical relationships is switched according to the processes of idling, contact establishment, stable drilling, drill string lifting, drill string running down, and rod connection / unloading; the set of physical relationships includes at least drilling speed consistency relationships and no-load torque consistency relationships, and the corresponding physical relationship residuals are determined according to the following formulas:

[0033]

[0034]

[0035] in, For drilling speed consistency residuals, This refers to the residual for consistency of no-load torque. For drilling speed, For drilling displacement, For time, For actual measured torque, For rotational speed, To match the rotational speed The corresponding no-load torque baseline;

[0036] Under stable drilling operations, the set of physical relationships also includes the consistency relationship between the ratio of propulsion force to torque and the theoretical range of mechanical specific energy, as well as the consistency relationship between single-rotation penetration depth and propulsion force; under drilling or drilling rig lifting operations, the set of physical relationships also includes the consistency relationship between displacement direction and operation state.

[0037] Furthermore, the smoothing confidence level is updated according to the following formula:

[0038]

[0039] in, For the confidence response coefficient, for the explicit failure state represented by continuous disconnection, range out-of-bounds, saturation or invalid status codes, skip time smoothing and directly reduce the smoothing confidence of the corresponding channel to below the failure threshold or set it to zero.

[0040] Furthermore, the channel information gain The redundancy of a channel is determined by the mutual information between channel characteristics and lithological categories, rock layer interface labels, or rock mechanical parameters, the importance of model permutation, or the validation error increment after deleting the channel; Adopting the first The weighted average of the correlation coefficient, mutual information, or conditional correlation between the path channel and other effective channels is determined; when the sum of the credibility of all channels is lower than the set threshold, the normalization weighting of the credibility of each channel is stopped, and the system directly enters the low-credibility output or stops the deterministic identification state.

[0041] Furthermore, the cross-sensor reconstruction is achieved through a pre-trained cross-sensor reconstruction model, the expression of which is as follows:

[0042]

[0043] It is based on the current drilling process Available channel data excluding abnormal channels The historical sequence of the abnormal channel before the anomaly occurred. and drilling rig control parameters Input: Abnormal channel estimate Reconstructing uncertainty and refactoring applicability Based on the reconstruction applicability, a reconstruction flag is generated to indicate whether the channel data is a reconstruction value;

[0044] when Higher than the reconstruction uncertainty threshold When the value is below the reconstruction applicability threshold or the current sample exceeds the applicable range of the reconstruction model, the reconstruction value will not be used to participate in the deterministic rock mass information identification, and the corresponding dynamic fusion weight will be reset to zero.

[0045] For completely failed channels, invoke the missing modality identification model that matches the combination of currently available channels;

[0046] The cross-sensor reconstruction also includes calculating a linear baseline reconstruction value for short-time missing channels, the linear baseline reconstruction value being calculated according to the following formula:

[0047]

[0048] in, Let be the average value of the i-th sensor under healthy operating conditions. Let be the mean vector of all channels except the i-th sensor under healthy operating conditions; This is the vector of current measurement values ​​for all channels except the i-th sensor. Let be the inverse covariance matrix among the channels other than the i-th sensor under healthy operating conditions. Let be the covariance vector of the i-th sensor and other channels under healthy operating conditions;

[0049] The reconstruction uncertainty is calculated according to the following formula:

[0050]

[0051] in, The variance of the cross-sensor reconstruction model's prediction of the current input and output. To reconstruct the local sensitivity matrix of the available channels for the cross-sensor model, This is the covariance matrix of the available channels under healthy operating conditions.

[0052] Furthermore, the sample distribution shift Determined using the Mahalanobis distance of the current fused features relative to the healthy training distribution:

[0053]

[0054] in, and Let be the mean vector and covariance matrix of the health training features, respectively; when When the set threshold is exceeded, the current state is marked as insufficient model applicability, the reliability of the results is reduced, and the output of reconstructed values ​​and deterministic categories is limited.

[0055] Furthermore, when the credibility of the results... When the confidence level is above the first threshold, output a deterministic identification result and its probability; when the confidence level is between the first threshold and the second threshold, output the identification result, low confidence marker and manual review prompt; when the confidence level is below the second threshold, stop outputting the deterministic category, switch to missing modality identification or alarm state, and output at least one of the following: channel, physical relationship residual, reconstruction uncertainty and sample distribution offset that caused the decrease in the confidence level of the result.

[0056] Furthermore, the missing modality recognition model is obtained by implementing random channel masking, channel noise perturbation, and modal distillation during the training of the rock mass information recognition model, or by training separately for predefined combinations of available channels. When calling the missing modality recognition model, the available channel mask, the smoothness confidence of each channel, the reconstruction flag, and the current drilling process are jointly input into the missing modality recognition model to distinguish between measured values, reconstructed values, and missing values.

[0057] This invention also discloses a fault-tolerant identification system for rock mass information based on dynamic weighting of multi-source drilling parameter reliability, comprising a synchronous acquisition module, an edge processor, and an output terminal; the synchronous acquisition module is used to acquire thrust, torque, rotational speed, drilling displacement, and drilling speed during the drilling process, and sends the acquired data to the edge processor; the edge processor includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method described in any one of claims 1 to 9; the output terminal is connected to the edge processor and is used to display the identification results and channel health status.

[0058] The beneficial effects of this invention are:

[0059] 1. Before multi-source data enters the rock mass information identification model, this invention explicitly calculates the dynamic reliability of each sensor from four dimensions: data integrity, signal stability, multi-parameter physical consistency, and historical health status, and performs time smoothing on it. This accurately distinguishes between two different situations: "low feature contribution" and "unreliable sensor measurement," avoiding the use of erroneous data as normal input in subsequent identification. Based on this, for six types of abnormal states—isolated jumps, short-term missing values, continuous drift, stuck values, saturation, and complete failure—differentiated fault-tolerant processing is performed, including outlier repair, cross-sensor reconstruction, baseline correction, reduction of dynamic reliability, or channel shielding. This overcomes the limitation of existing single filtering methods in handling complex anomaly types, maintaining the continuous availability of the identification system even under partial sensor failure conditions.

[0060] 2. This invention constructs a set of physical relationships based on different drilling processes, such as idling, contact establishment, stabilizing drilling, drill bit lifting, drill bit running down, and rod unloading. This allows the same sensor signal to be evaluated using different rationality relationships under different processes. For example, a reverse change in displacement during drill bit lifting is normal, while a sudden reverse during stabilizing drilling is considered abnormal. This process-dependent evaluation mechanism effectively avoids misjudging normal process switching as sensor failure, and prevents missed detections due to overly wide thresholds when a real fault occurs, significantly improving the accuracy of anomaly identification.

[0061] 3. In the cross-sensor reconstruction process, this invention not only outputs estimated values ​​of abnormal channels but also simultaneously outputs reconstruction uncertainty and reconstruction applicability. When the reconstruction uncertainty exceeds a threshold, the reconstruction applicability falls below a threshold, or the current sample exceeds the applicable range of the reconstruction model, the reconstructed value is not used in deterministic rock mass information identification, and the corresponding dynamic fusion weight is reset to zero. This mechanism effectively avoids the adoption of reconstructed values ​​that appear smooth but have physical errors by the subsequent identification model, thereby preventing high-confidence misjudgments caused by erroneous reconstruction. Simultaneously, the linear benchmark reconstruction value provides a physical rationality verification basis for the nonlinear reconstruction results, further ensuring the reliability of the reconstruction results.

[0062] 4. This invention considers three factors simultaneously when calculating dynamic fusion weights: smoothness confidence, channel information gain, and channel redundancy. Information gain ensures that channels contributing significantly to the current recognition task receive higher weights, while the redundancy penalty term prevents multiple highly correlated channels from repeatedly amplifying the same error. It also preserves supplementary information from independent modes such as vibration regarding cracks, cavities, and abnormal drilling. This multi-dimensional weighting mechanism is superior to traditional methods that only normalize weights based on confidence, enabling fusion features to fully utilize effective information while suppressing the propagation of redundant errors.

[0063] 5. This invention comprehensively considers six dimensions—prediction certainty, effective channel availability, overall physical consistency of fused data, reconstruction uncertainty, sample distribution offset, and effective channel redundancy—to calculate the reliability of the results. Based on the reliability, a three-tiered output strategy is implemented: when the reliability exceeds a first threshold, a deterministic identification result and its probability are output; when the reliability is between the first and second thresholds, the identification result, a low-confidence marker, and a manual review prompt are output; when the reliability is below the second threshold, the output of deterministic categories is stopped, and the system transitions to missing mode identification or an alarm state. This gating mechanism provides a traceable quality control basis for on-site drilling lithology identification, stratum interface location, and rock strength prediction, effectively reducing the risk of high-confidence misjudgments caused by erroneous data.

[0064] 6. This invention trains the missing mode identification model through methods such as random channel masking, channel noise perturbation, and modal distillation, or trains independent models for each predefined combination of available channels. This allows the system to still identify rock mass information using the remaining channels after completely failed channels have been removed. When calling the missing mode identification model, the available channel mask, the smoothness confidence of each channel, the reconstruction flag, and the current drilling process are input into the model, enabling the model to distinguish between measured values, reconstructed values, and missing values, further improving the system's degraded operation capability under extreme failure conditions.

[0065] 7. The fault-tolerant processing of this invention forms a traceable record. Each sliding window records at least the original channel state, anomaly type, reliability, repair or reconstruction method, reconstruction uncertainty, final fusion weight, and output reliability. These records are not only used for on-site verification and sensor maintenance, but also serve as training data for subsequent model updates. This avoids the drawback of only saving the final lithology results without being able to trace data quality, and provides a data foundation for continuous system optimization. Attached Figure Description

[0066] Figure 1 This is a schematic diagram of the sensor reliability evaluation, anomaly handling, and dynamic weighting structure.

[0067] Figure 2 A flowchart for fault-tolerant identification of rock mass information. Detailed Implementation

[0068] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0069] Example 1: The rock mass information fault-tolerant identification method based on dynamic weighting of multi-source drilling parameter reliability in this example includes the following steps:

[0070] 1) Synchronous acquisition of multi-source drilling parameters and identification of drilling procedures

[0071] Thrusting force sensors, torque sensors, rotational speed sensors, displacement sensors, and velocity sensors are deployed on the drilling rig to collect thrust, torque, rotational speed, drilling displacement, and drilling speed during the drilling process. All sensors are synchronously sampled using a unified clock source with a sampling frequency of 100 Hz. After aligning the data from each channel according to a unified timestamp, a sliding window segment with a window length of 5 seconds and a window overlap rate of 50% is formed, creating the k-th multi-source sliding window X. k Each sliding window contains 500 sampling points, and adjacent windows share 2.5 seconds of data. The drilling speed can be calculated from the displacement difference or directly acquired by an independent velocity sensor; as a preferred option, this embodiment uses an independent velocity sensor for direct acquisition to cross-validate with the displacement difference value and improve the accuracy of the physical consistency evaluation.

[0072] Within each sliding window, the current drilling operation is identified based on the thrust, torque, rotational speed, and the direction and duration of the drilling displacement or drilling speed. k .

[0073] In this embodiment, the judgment rules for each process are as follows:

[0074] Idle operation: For example, when the thrust is less than 5 kN, the torque is less than 200 N·m, the speed is positive and stable within the range of 20-60 rpm with fluctuations of less than 5%, and the displacement change within a 5-second window is less than 2 mm, it is judged as idle operation.

[0075] Contact establishment process: For example, when the thrust gradually increases from the no-load level to about 30 kN, the torque increases synchronously, the speed decreases by no more than 10%, and the displacement begins to increase positively, it is determined that contact has been established.

[0076] Stable drilling process: For example, when the thrust is maintained at 30-80 kN, the torque is stable at 500-1500 N·m, the rotational speed is maintained at 20-40 rpm, the displacement continues to increase in the positive direction, and the drilling speed is within the range of 0.1-0.5 m / min, it is judged as stable drilling.

[0077] Drilling lifting procedure: For example, when the thrust drops to the no-load level, the torque drops to the no-load level, and the displacement direction changes from positive to negative, it is determined to be drilling lifting.

[0078] Drilling process: For example, when the thrust is kept unloaded or slightly positive, the torque is kept unloaded, the displacement direction is positive and the speed is greater than 0.5 m / min, it is determined to be drilling.

[0079] Rod connection / disconnection process: For example, when the thrust drops to no load, the torque drops to no load, the speed drops to zero and then recovers, and the displacement change is less than 1 mm within 5 seconds, it is determined to be rod connection / disconnection.

[0080] During the process identification, the direction of the power head's movement is determined based on the direction of the drilling displacement or drilling speed.

[0081] It should be noted that the above values ​​are all exemplary values ​​in this embodiment. In actual application, they can be adjusted according to the drilling rig model, geological conditions and engineering experience. Those skilled in the art can understand that the changes in these values ​​do not depart from the protection scope of this invention.

[0082] 2) Four-dimensional credibility evaluation and smoothing

[0083] For the i-th sensor, calculate the data integrity score respectively. Signal stability score Multi-parameter physical consistency scoring and historical health score .

[0084] Data integrity score Based on the missing rate of the i-th sensor within the current sliding window. Packet loss rate Range out-of-bounds rate saturation rate and continuous card value ratio Calculate using the following formula:

[0085]

[0086] in, to The weights are non-negative; in this embodiment, to The values ​​are 0.3, 0.2, 0.2, 0.2, and 0.1 respectively. The missing rate is determined by the proportion of null or invalid values, the communication packet loss rate is determined by the proportion of discontinuous acquisition sequence numbers, the range out-of-range rate is determined by the proportion of measured values ​​exceeding the calibrated range, the saturation rate is determined by the proportion of measured values ​​continuously approaching the upper or lower limit of the range, and the continuous card value ratio is determined by the proportion of measured values ​​remaining completely unchanged within a set time.

[0087] Signal stability score Based on the local noise ratio of the i-th sensor within the current sliding window , jump rate Drift slope Periodic interference energy and the proportion of abnormal peak values Calculate using the following formula:

[0088]

[0089] in, to The weights are non-negative; in this embodiment, to The values ​​were 0.4, 0.3, 0.2, 0.05, and 0.05, respectively. The median and median absolute deviation of the most recent 20 healthy windows were used as baselines. The local noise ratio was determined by the ratio of high-frequency residual energy to effective signal energy. For the vibration channel, the high-frequency residual energy was calculated from the residual obtained by subtracting the 0.1-second window mean from the 5 kHz original vibration signal, effectively reflecting the noise level of the vibration signal. The jump rate was determined by the proportion of sampling points whose adjacent differences exceeded the median absolute deviation threshold, and the drift slope was determined by the rate of change of the current window baseline relative to the historical healthy baseline. For the Hall rotation speed pulse signal, if the coefficient of variation of the pulse interval exceeded twice the healthy baseline, a periodic interference term was added.

[0090] Multi-parameter physical consistency score The residuals between the physical relationships of the drilling parameters corresponding to the i-th sensor are calculated using the following formula:

[0091]

[0092] in, For the current drilling process, including the first The set of physical relationships of road sensors. For the corresponding physical relationship residuals, For health standards, Non-negative weights are used. The set of physical relationships switches according to the process.

[0093] In this embodiment, the physical relationship residuals include, but are not limited to, the following:

[0094] Drilling speed consistency residual Calculate according to the following formula:

[0095]

[0096] in, For drilling speed, For drilling displacement, For time, For the measured torque, this residual is used to verify the consistency between the independent velocity sensor measurement value and the displacement differential calculation value;

[0097] No-load torque consistency residual Calculate according to the following formula:

[0098]

[0099] in, For actual measured torque, For rotational speed, To match the rotational speed The corresponding no-load torque baseline; this residual is used to verify the consistency between the torque sensor's measurement value under no-load conditions and the historical baseline.

[0100] thrust-torque ratio consistency relationship residual Under stable drilling conditions, the residual is defined as the deviation between the ratio of actual thrust to torque and the theoretical expected value of mechanical specific energy, calculated according to the following formula:

[0101]

[0102] in, For propulsion, For torque, This represents the theoretical ratio of mechanical specific energy under the current process. If this residual exceeds the health threshold, it indicates that the propulsion or torque sensor may be malfunctioning.

[0103] Consistency between displacement direction and process state residual During the drilling or hoisting process, this residual is a logical residual and is calculated according to the following formula:

[0104]

[0105] Specifically, the displacement is consistent in the reverse direction when the drill string is pulled up, and consistent in the forward direction when the drill string is lowered. If the residual value is 1, it is considered abnormal.

[0106] Displacement stability residual in idling state Under idling conditions, the residual is defined as the absolute value of the rate of change of displacement, calculated according to the following formula:

[0107]

[0108] If the residual exceeds the health scale, it indicates that the displacement sensor may be drifting.

[0109] Speed ​​pulse consistency residual The residual is defined as the deviation between the output value of the speed sensor and the converted value of the Hall pulse count per unit time, and is calculated according to the following formula:

[0110]

[0111] in, The output value of the speed sensor. This is the count of Hall pulses within a unit time Δt. This represents the number of pulses per revolution. If this residual exceeds the health threshold, it indicates a potential malfunction in the speed sensor.

[0112] Each residual is normalized according to the health scale parameters and then weighted and summed to obtain a multi-parameter physical consistency score.

[0113] To further quantify the deviation of each channel from the overall physical laws, this embodiment constructs a graph structure composed of sensor channels and physical relationships, and introduces... Figure 1 Anomaly detection is performed on the consistency residuals. The i-th channel's... Figure 1 Consistency residuals are obtained by weighted normalization of the associated physical relationship residuals, and the calculation formula is as follows:

[0114]

[0115] Among them, g k (x) represents the residual of the k-th physical relation, N i Let β be the set of physical relationships containing the i-th channel. ik δ represents the relation weight. k As a health standard; Figure 1 The larger the consistency residual, the more inconsistent the channel is with the overall physical laws, and the lower its physical consistency score.

[0116] Set of physical relationships under different drilling procedures The composition of these relationships varies. For example, the physical relationship set for the idling process mainly includes the consistency relationship of no-load torque and the consistency relationship of rotational speed pulses; the stable drilling process also includes the consistency relationship of thrust-torque ratio, drilling speed, and mechanical specific energy; and the drilling rig lifting or lowering process also includes the consistency relationship between displacement direction and process state. The physical relationship set corresponding to each process can be pre-configured based on the drilling rig model and engineering experience.

[0117] Historical health score Initialized from equipment maintenance records. Newly installed sensors that have undergone zero-point, full-scale, and repeatability calibration. Take 0.95; gradually reduce after the cumulative running time reaches 80% of the maintenance cycle. When drift alarms occur three times in a row, Decrease by 0.1; restore to 0.95 after successful recalibration. The initial method for establishing the health baseline is as follows: During the equipment commissioning phase or the initial stage of system operation, collect data for N consecutive health windows (e.g., N=50), and calculate the mean vector of each channel within that time period. Covariance Matrix This serves as the initial health baseline. During subsequent operation, the health baseline is dynamically updated using an exponentially weighted moving average method. Data within fault windows, process switching windows, or manually marked abnormal windows are masked and do not participate in the baseline update calculation.

[0118] Substitute the above four-dimensional scores into the following formula to calculate the dynamic reliability of the i-th sensor. :

[0119]

[0120] in, The input is a function that maps to the interval between 0 and 1. a, b, c, d, and e are preset parameters or parameters determined by calibration data. In this embodiment, σ is the Sigmoid function, a, b, c, and d are 1.2, 1.0, 1.5, and 0.5 respectively, and e is -2.0.

[0121] The dynamic confidence level is smoothed over time to obtain the smoothed confidence level. :

[0122]

[0123] in, The credibility response coefficient is used in this embodiment. Set the value to 0.3. For explicit failure states represented by continuous disconnection, range out-of-range, saturation, or invalid status codes, skip time smoothing and directly reduce the smoothing reliability of the corresponding channel to below the failure threshold or set it to zero.

[0124] A dynamic reliability score greater than 0.8 is considered healthy, 0.5 to 0.8 is considered slightly abnormal, 0.2 to 0.5 is considered seriously abnormal, and less than 0.2 is considered invalid.

[0125] 3) Identify abnormal states based on the magnitude of the smoothing confidence level and its changing trend and rate of change between adjacent sliding windows, and confirm the anomaly type based on the duration of the anomaly. Specific determination rules are as follows:

[0126] Isolated jump: The smooth confidence level suddenly drops by more than 0.3 within a single window, but the confidence level of the preceding and following windows returns to normal, and the duration of the anomaly does not exceed one window.

[0127] Short-term missing data: Data integrity score is below 0.5 and the duration does not exceed 5 windows (25 seconds).

[0128] Continuous drift: The smoothing confidence level decreases unidirectionally over more than 5 consecutive windows, and the physical consistency score decreases synchronously.

[0129] Card value: The signal stability score is below 0.3, the signal variance is close to zero, and there are obvious changes in related processes or other channels.

[0130] Saturation: The signal continues to approach the range boundary and the derivative is truncated for more than 3 windows.

[0131] Complete failure: Smoothing confidence is below 0.2 and lasts for more than 5 windows, or continuous missing for more than 5 seconds.

[0132] Perform fault tolerance processing for the aforementioned abnormal state:

[0133] Isolated abrupt changes: Hampel filtering is used for repair. Taking the outlier as the center, five sampling points are taken before and after it, and the median and median absolute deviation are calculated. Sampling points that exceed the median ± 3 times the median absolute deviation are replaced with the median.

[0134] Short-term missing: Perform cross-sensor reconstruction. Cross-sensor reconstruction is achieved through a pre-trained cross-sensor reconstruction model, the expression of which is as follows:

[0135]

[0136] In this embodiment, the current drilling process is used. Available channel data excluding abnormal channels The historical sequence of the abnormal channel 2 seconds before the anomaly occurred. and drilling rig control parameters As input to the cross-sensor reconstruction model, the output is anomaly channel estimates. Reconstructing uncertainty Reconstruction Applicability In this embodiment, the cross-sensor reconstruction model can employ various machine learning models, including but not limited to the gradient boosting model (LightGBM), random forest, long short-term memory network, or combinations thereof; as a preferred option, this embodiment employs the gradient boosting model (LightGBM).

[0137] when The value exceeds the reconstruction uncertainty threshold (in this embodiment, it is set to be twice the healthy torque variance). If the value is below the reconstruction applicability threshold (set to 0.6 in this embodiment) or the current sample exceeds the applicable range of the reconstruction model, the reconstruction value will not be used for deterministic rock mass information identification, and the corresponding dynamic fusion weight will be reset to zero. If the current channel is temporarily missing and If the value is not lower than 0.6, a reconstruction flag is generated. =1 indicates that the current data is a reconstructed value, and will... The current measurement value of this channel is used in subsequent processing. The reconstruction flag is a multi-state flag, defined in this embodiment as follows: =0 indicates that the data in this channel is the measured value. =1 indicates that the channel data is a reconstructed value. =2 indicates that the data in this channel is missing.

[0138] Cross-sensor reconstruction also includes calculating linear baseline reconstruction values ​​for short-time missing channels, serving as a basis for verifying the physical rationality of the nonlinear reconstruction results. The linear baseline reconstruction value is calculated using the following formula:

[0139]

[0140] in, Let be the average value of the i-th sensor under healthy operating conditions. Let be the mean vector of all channels except the i-th sensor under healthy operating conditions; This is the vector of current measurement values ​​for all channels except the i-th sensor. Let be the inverse covariance matrix among the channels other than the i-th sensor under healthy operating conditions. Let be the covariance vector of the i-th sensor and other channels under healthy operating conditions;

[0141] The reconstruction uncertainty is calculated according to the following formula:

[0142]

[0143] in, The variance of the cross-sensor reconstruction model's prediction of the current input and output is (estimated by the leaf node variance of the gradient boosting model). The local sensitivity matrix of the cross-sensor reconstruction model for the available channels is calculated using the finite difference method. This is the covariance matrix of the available channels under healthy operating conditions.

[0144] Continuous drift: Corrected based on the idle baseline. This embodiment uses an exponentially weighted moving average to dynamically update the health baseline to adapt to the slow drift of equipment operating status. Data within fault windows, process switching windows, or manually marked abnormal windows are masked and not included in the baseline update calculation, thus ensuring that the health baseline always reflects the normal degradation trend of the equipment. The median deviation of the most recent idle period is used as the drift amount, which is subtracted from the current measurement value.

[0145] Card value and saturation: Reduce the dynamic confidence level of the corresponding channel to below 0.3, or directly block the channel (set the dynamic confidence level to zero).

[0146] Complete Failure: The missing modality recognition model is obtained by implementing random channel masking (randomly masking 30% of channels during each training session), channel noise perturbation (adding Gaussian noise with a standard deviation of 10% of the healthy signal's standard deviation), and modal distillation during the training of the rock mass information recognition model. Modal distillation is implemented as follows: using the complete modality model as the teacher network and the missing modality model as the student network, the student network learns to simulate the teacher network's output distribution even with some channels missing by minimizing the KL divergence between the teacher and student network outputs. When calling the missing modality recognition model, the available channel mask, the smoothness confidence of each channel, the reconstruction flag indicating the data source of each channel, and the current drilling operation are input into the model, enabling the model to distinguish between measured values, reconstructed values, and missing values.

[0147] 4) Dynamic weighting and feature fusion

[0148] Extract the features of each channel after fault tolerance processing. In this embodiment, the following features are extracted for each low-frequency signal (propulsion, torque, rotational speed, displacement, velocity): median, standard deviation, rate of change, frequency band energy (0-10Hz, 10-50Hz, 50-100Hz), mechanical specific energy related features (torque × rotational speed / propulsion / drilling speed), and temporal embedding features (LSTM embedding vector extracted using a sliding window of length 10).

[0149] Based on smoothness reliability Channel information gain and channel redundancy The final dynamic fusion weight is calculated using the following formula. :

[0150]

[0151] in, and These are the channel information gain and channel redundancy of the j-th channel, respectively. Number of sensor channels , and This is the adjustment coefficient; in this embodiment, Take 1.0, Take 0.5, Set to 0.3. Channel information gain. The redundancy is determined using the mutual information between channel characteristics and lithological categories, obtained during the offline calibration phase, and can be updated periodically based on field-confirmed samples. The correlation coefficient between the i-th channel and other effective channels is used to determine the correlation.

[0152] Constructing fusion features For channels that are removed from the algorithm, their dynamic fusion weight is zero; for channels that use reconstructed values, their dynamic fusion weight does not exceed 0.2. At the same time, a maximum weight limit of 0.4 is set for a single channel to prevent a high information gain channel from still occupying an excessively high weight when its credibility slightly decreases.

[0153] 5) Rock mass information identification and result reliability gating

[0154] The fusion features The input is a pre-trained rock mass information recognition model, which can employ random forest, gradient boosting model, long short-term memory network, Transformer, or a combination thereof. In this embodiment, the rock mass information recognition model uses a temporal classification model, consisting of two LSTM layers and one fully connected layer, outputting the probabilities of four lithology categories: limestone, sandstone, mudstone, and shale, as well as the prediction entropy. During the training phase, the model uses historical health data for supervised learning, with lithology category as the label. When outputting the recognition results, the depth of the center of the sliding window is used as the depth corresponding to the category of the current window; when the category probabilities of adjacent windows continuously change, the midpoint of the probability change is determined as the location of the rock layer interface and output.

[0155] The reliability of the result is calculated according to the following formula. :

[0156]

[0157] in, To be based on the predicted entropy Certainty of prediction , The maximum entropy corresponds to the four categories.

[0158] The effective channel confidence is weighted and summed, and the weighted average of all channels with a smooth confidence greater than 0.5 is taken.

[0159] To ensure overall physical consistency of the data, it is determined by a weighted average of the residuals of all valid physical relationships.

[0160] To reconstruct uncertainty, take the maximum normalized reconstruction uncertainty of all channels using reconstruction values.

[0161] The sample distribution shift is determined using the Mahalanobis distance between the current fused features and the healthy training features:

[0162]

[0163] in, and These are the mean vector and covariance matrix of the health training features, respectively; the health training features are derived from fused feature data under historical health conditions. When the threshold is exceeded (set to 0.3 in this embodiment), the current state is marked as insufficient model applicability, the reliability of the results is reduced and the output of reconstructed values ​​and deterministic categories is restricted.

[0164] The effective channel redundancy is determined by the average correlation coefficient between effective channels.

[0165] to As weights, in this embodiment to Take values ​​of 0.4, 0.3, 0.1, 0.1, 0.05, and 0.05 respectively.

[0166] `clip` means to restrict the result to the range of 0 to 1.

[0167] Three output levels are executed based on the reliability of the results:

[0168] Normal output: When When the value is ≥ 0.75, output the deterministic lithology category and its probability, and output the location of the rock layer interface (the depth corresponding to the continuous change of the category probability of adjacent windows).

[0169] Low-confidence review warning: When 0.45 ≤ When the value is less than 0.75, the identification result is output and marked "low confidence, it is recommended to verify with core samples or adjacent wells".

[0170] Stop deterministic identification: when When < 0.45, no deterministic lithology category is output, and the system switches to missing mode identification or alarm state, and outputs the main reasons for the reduced reliability of the results, including at least one of abnormal channels, physical relationship residuals, reconstruction uncertainty and sample distribution shift.

[0171] In addition to displaying lithology type or intensity prediction, the output terminal also displays the channel health matrix for each sensor channel. ,point , , , and finally The processing records, such as "torque drift correction", "displacement short-term loss reconstruction" and "speed channel rejection", are listed in text.

[0172] Example 2: Based on Example 1 above, this example further adopts the following preferred solution:

[0173] The preferred solution in step 1: In this embodiment, vibration signals and motor current signals are also collected simultaneously. The vibration sensor is a triaxial accelerometer with a sampling frequency of 5 kHz, used to capture the high-frequency vibration characteristics generated by the interaction between the drill bit and the rock mass; the motor current sensor is installed on the power line of the drive motor, with a sampling frequency consistent with low-frequency parameters such as thrust, set to 100 Hz, used to monitor the load changes of the drive motor.

[0174] The preferred solution in step 2: In this embodiment, the motor current signal, torque, and speed are all used to verify the physical consistency of the transmission system. Under stable drilling conditions, the motor current and torque are positively correlated. When the torque increases but the motor current does not increase synchronously, it is determined that the torque sensor may be drifting; conversely, when the motor current increases but the torque does not increase synchronously, it is determined that the motor current sensor may be malfunctioning.

[0175] The preferred solution in step 3: In this embodiment, vibration signals are used to help distinguish between sensor malfunctions and real geological events during the anomaly identification process. When torque, thrust, and vibration signals simultaneously experience impact, the system retains this segment as a real anomaly event (such as the drill bit encountering a hard interlayer or fracture), only lowering the upper limit of measurement reliability, and does not delete it directly; when only torque or thrust is abnormal while the vibration signal is stable, it is determined to be a sensor malfunction, and corresponding fault-tolerant processing is performed.

[0176] The preferred solution in step 4: In this embodiment, for vibration signals, the root mean square, kurtosis, and energy of each frequency band (0-500 Hz, 500-2000 Hz, 2000-5000 Hz) within a 0.1-second window are extracted as supplementary features, which participate in dynamic fusion together with low-frequency signal features.

[0177] Example 3: A fault-tolerant identification system for rock mass information based on dynamic weighting of multi-source drilling parameter reliability, including a synchronous acquisition module, an edge processor, and an output terminal.

[0178] The synchronous acquisition module is used to collect the thrust, torque, rotational speed, drilling displacement and drilling speed during the drilling process, and send the collected data to the edge processor.

[0179] The edge processor employs an embedded industrial control computer or an edge computing gateway, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described in Embodiment 1 or Embodiment 2. Specifically, after receiving multi-source drilling parameter data from the synchronous acquisition module, the edge processor sequentially performs steps such as drilling process identification, four-dimensional reliability evaluation and smoothing, anomaly identification and fault tolerance processing, dynamic weighting and feature fusion, rock mass information identification and result reliability gating, generating identification results such as lithology category, rock layer interface location, and rock mechanical parameters, as well as health status information for each channel.

[0180] The output terminal uses an industrial touchscreen or industrial control display, connected to the edge processor, to display the identification results and channel health status. Specifically, the output terminal displays key information such as lithology type and its probability, rock layer interface location, and result reliability in real time through a graphical interface. It also displays the data integrity score, signal stability score, multi-parameter physical consistency score, historical health score, dynamic reliability, and final dynamic fusion weight of each sensor in the form of a channel health matrix. At the same time, it lists fault-tolerant processing records such as "torque drift correction," "displacement short-term missing reconstruction," and "speed channel rejection" in text.

[0181] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A fault-tolerant identification method for rock mass information based on dynamic weighting of multi-source drilling parameter reliability, characterized in that: Includes the following steps: 1) Synchronously collect thrust, torque, rotational speed, drilling displacement, and drilling speed during the drilling process, and form the k-th multi-source sliding window X according to a unified time reference. k It identifies the current drilling operation based on the thrust, torque, rotational speed, and the direction and duration of drilling displacement or drilling speed. k ; 2) For the i-th sensor, calculate the data integrity score respectively. Signal stability score Multi-parameter physical consistency scoring and historical health score Among them, the multi-parameter physical consistency score According to the current drilling procedure s k Calculate the physical relationship residuals between the following drilling parameters; calculate the dynamic reliability of the i-th sensor according to the following formula. : in, For a function that maps the input to the interval between 0 and 1, a, b, c, d, and e are preset parameters or parameters determined by calibration data; The dynamic confidence level is smoothed over time to obtain the smoothed confidence level. ; 3) Identify abnormal states based on the magnitude of the smoothing confidence level and its changing trend and rate of change between adjacent sliding windows. Confirm isolated jumps, short-term missing values, persistent drift, stuck values, saturation, and complete failure states based on the duration of the anomaly. Perform fault-tolerant processing on these abnormal states, including: outlier repair for isolated jumps, cross-sensor reconstruction for short-term missing values, baseline correction for persistent drift, reduction of dynamic confidence level or channel masking for stuck values ​​and saturation, and channel masking for complete failure. The cross-sensor reconstruction outputs at least the estimated abnormal channel value, reconstruction uncertainty, and reconstruction applicability. 4) Extract the features of each channel after fault tolerance processing. Based on smoothness confidence Channel information gain and channel redundancy The final dynamic fusion weight is calculated using the following formula. : in, and These are the channel information gain and channel redundancy of the j-th channel, respectively. Number of sensor channels , and This is the adjustment coefficient; Constructing fusion features ; 5) The fusion features Input a pre-trained rock mass information recognition model, which is used to output at least one of the following recognition results based on fused features: lithology category, rock layer interface location, and rock mechanical parameters, as well as prediction entropy, and calculate the reliability of the result according to the following formula. : in, The degree of prediction certainty is determined based on the prediction entropy. Weighted sum of the credibility of effective channels, To ensure overall physical consistency of the data, To reconstruct uncertainty, For sample distribution shift, For effective channel redundancy; to The weight is ; clip indicates that the result is restricted to the range of 0 to 1; based on the confidence level of the result, normal output, low confidence review prompt, or stop deterministic identification is performed.

2. The rock mass information fault-tolerant identification method based on dynamic weighting of multi-source drilling parameter reliability according to claim 1, characterized in that: The data integrity score Based on the missing rate of the i-th sensor within the current sliding window Packet loss rate Range out-of-bounds rate saturation rate and continuous card value ratio Calculate using the following formula: in, to Non-negative weights; The signal stability score Based on the local noise ratio of the i-th sensor within the current sliding window , jump rate Drift slope Periodic interference energy and the proportion of abnormal peak values Calculate using the following formula: in, to Non-negative weights; The multi-parameter physical consistency score The residuals between the physical relationships of the drilling parameters corresponding to the i-th sensor are calculated using the following formula: in, For the current drilling process, including the first The set of physical relationships of road sensors. For the corresponding physical relationship residuals, For health standards, The weights are non-negative.

3. The rock mass information fault-tolerant identification method based on dynamic weighting of multi-source drilling parameter reliability according to claim 2, characterized in that: The set of physical relationships is switched according to the processes of idling, contact establishment, stable drilling, drill string lifting, drill string running, and rod connection / unloading; the set of physical relationships includes at least drilling speed consistency relationships and no-load torque consistency relationships, and the corresponding physical relationship residuals are determined according to the following formulas: in, For drilling speed consistency residuals, This refers to the residual for consistency of no-load torque. For drilling speed, For drilling displacement, For time, For actual measured torque, For rotational speed, To match the rotational speed The corresponding no-load torque baseline; Under stable drilling operations, the set of physical relationships also includes the consistency relationship between the ratio of propulsion force to torque and the theoretical range of mechanical specific energy, as well as the consistency relationship between single-rotation penetration depth and propulsion force; under drilling or drilling rig lifting operations, the set of physical relationships also includes the consistency relationship between displacement direction and operation state.

4. The rock mass information fault-tolerant identification method based on dynamic weighting of multi-source drilling parameter reliability according to claim 1, characterized in that: The smoothness confidence level is updated according to the following formula: in, For the confidence response coefficient, for the explicit failure state represented by continuous disconnection, range out-of-bounds, saturation or invalid status codes, skip time smoothing and directly reduce the smoothing confidence of the corresponding channel to below the failure threshold or set it to zero.

5. The rock mass information fault-tolerant identification method based on dynamic weighting of multi-source drilling parameter reliability according to claim 1, characterized in that: The channel information gain The redundancy of a channel is determined by the mutual information between channel characteristics and lithological categories, rock layer interface labels, or rock mechanical parameters, the importance of model permutation, or the validation error increment after deleting the channel; Adopting the first The weighted average of the correlation coefficient, mutual information, or conditional correlation between the path channel and other effective channels is determined; when the sum of the credibility of all channels is lower than the set threshold, the normalization weighting of the credibility of each channel is stopped, and the system directly enters the low-credibility output or stops the deterministic identification state.

6. The rock mass information fault-tolerant identification method based on dynamic weighting of multi-source drilling parameter reliability according to claim 1, characterized in that: The cross-sensor reconstruction is achieved through a pre-trained cross-sensor reconstruction model, the expression of which is as follows: It is based on the current drilling process Available channel data excluding abnormal channels The historical sequence of the abnormal channel before the anomaly occurred. and drilling rig control parameters Input: Abnormal channel estimate Reconstructing uncertainty and the applicability of reconstruction Based on the reconstruction applicability, a reconstruction flag is generated to indicate whether the channel data is a reconstruction value; when Higher than the reconstruction uncertainty threshold When the value is below the reconstruction applicability threshold or the current sample exceeds the applicable range of the reconstruction model, the reconstruction value will not be used to participate in the deterministic rock mass information identification, and the corresponding dynamic fusion weight will be reset to zero. For completely failed channels, invoke the missing modality identification model that matches the combination of currently available channels; The cross-sensor reconstruction also includes calculating a linear baseline reconstruction value for short-time missing channels, the linear baseline reconstruction value being calculated according to the following formula: in, Let be the average value of the i-th sensor under healthy operating conditions. Let be the mean vector of all channels except the i-th sensor under healthy operating conditions; This is the vector of current measurement values ​​for all channels except the i-th sensor. Let be the inverse covariance matrix among the channels other than the i-th sensor under healthy operating conditions. Let be the covariance vector of the i-th sensor and other channels under healthy operating conditions; The reconstruction uncertainty is calculated according to the following formula: in, The variance of the cross-sensor reconstruction model's prediction of the current input and output. To reconstruct the local sensitivity matrix of the available channels for the cross-sensor model, This is the covariance matrix of the available channels under healthy operating conditions.

7. The rock mass information fault-tolerant identification method based on dynamic weighting of multi-source drilling parameter reliability according to claim 1, characterized in that: The sample distribution offset Determined using the Mahalanobis distance of the current fused features relative to the healthy training distribution: in, and Let be the mean vector and covariance matrix of the health training features, respectively; when When the set threshold is exceeded, the current state is marked as insufficient model applicability, the reliability of the results is reduced, and the output of reconstructed values ​​and deterministic categories is limited.

8. The rock mass information fault-tolerant identification method based on dynamic weighting of multi-source drilling parameter reliability according to claim 1, characterized in that: When the reliability of the results When the confidence level is above the first threshold, output a deterministic identification result and its probability; when the confidence level is between the first threshold and the second threshold, output the identification result, low confidence marker and manual review prompt; when the confidence level is below the second threshold, stop outputting the deterministic category, switch to missing modality identification or alarm state, and output at least one of the following: channel, physical relationship residual, reconstruction uncertainty and sample distribution offset that caused the decrease in the confidence level of the result.

9. The rock mass information fault-tolerant identification method based on dynamic weighting of multi-source drilling parameter reliability according to claim 6, characterized in that: The missing modality recognition model is obtained by implementing random channel masking, channel noise perturbation, and modal distillation during the training of the rock mass information recognition model, or by training separately for predefined available channel combinations. When calling the missing modality recognition model, the available channel mask, the smoothness confidence of each channel, the reconstruction flag, and the current drilling process are jointly input into the missing modality recognition model to distinguish between measured values, reconstructed values, and missing values.

10. A fault-tolerant identification system for rock mass information based on dynamic weighting of multi-source drilling parameter reliability, characterized in that: The system includes a synchronous acquisition module, an edge processor, and an output terminal. The synchronous acquisition module is used to acquire the thrust, torque, rotational speed, drilling displacement, and drilling speed during the drilling process, and sends the acquired data to the edge processor. The edge processor includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described in any one of claims 1 to 9. The output terminal is connected to the edge processor and is used to display the identification results and channel health status.

Images