Dynamic torque limiting and early warning linkage edge control method and system for top drive system

Abstract

The application discloses a kind of dynamic limit torsion and early warning linkage edge control method and system for top drive system, it is related to petroleum drilling equipment technical field, the application obtains target multi-source sensing data, eliminates the data barrier between different protocol sensors.Risk assessment is carried out on the time series formed by target multi-source sensing data using trained TCN-Transformer model, which can identify the risk of sticking and sticking in the early stage of weak signs.According to the probability of abnormal occurrence and the type of abnormal signs, combined with the rated performance parameters of the top drive system and the real-time working condition information, a dynamic torque limit curve is generated, which allows the upper limit of torque to be adjusted adaptively according to the risk level and formation conditions.When the probability of abnormal occurrence exceeds the preset threshold, a cooperative control signal is sent to the well control safety system to trigger the well control related devices to enter a preventive state in advance, realizing the multi-system parallel linkage of top drive restriction and well control protection.

Description

Technical Field

[0001] This invention relates to the field of oil drilling equipment technology, and in particular to a dynamic torque limiting and early warning linkage edge control method and system for top drive systems. Background Technology

[0002] In oil drilling operations, the top drive system is a key piece of equipment used to drive the drill string rotation and control drilling parameters. Its torque control strategy directly affects drilling safety and operational efficiency. With the increasing demand for drilling deep wells, ultra-deep wells, and complex formations, formation changes are frequent and drilling pressure fluctuations are severe during the drilling process. Early risk signs such as abnormal torque fluctuations, sudden increases in hook load, and sudden drops in drilling speed are prone to occur. If these risks are not detected in time and coordinated torque limiting and protection measures are not taken, it can easily lead to stuck pipe, drill string overload, or even equipment damage. Therefore, how to perceive multi-source heterogeneous data in real time during the drilling process, proactively identify weak abnormal signs, dynamically generate and execute torque limiting strategies, and coordinate with the well control system to implement protection has become a key technical issue in the current intelligent and safe development of top drive operations.

[0003] To address the aforementioned issues, patent CN119900457A discloses a top drive torque adjustment method and system. This method uses sensors deployed in the top drive system to obtain operating condition data such as drilling pressure, bottom hole rotation speed, and mud flow rate. An initial neural network model is constructed, trained using historical operating condition data as input and operating torque as output. The model predicts torque based on current operating conditions and adjusts the motor output torque based on real-time operating torque. Simultaneously, a frequency converter adjusts the frequency and voltage of the motor's input power supply to change the speed and torque output characteristics. The motor's operating temperature and status are monitored, and power is cut off or output torque is reduced when the motor is overloaded. Patent CN120487034A discloses a drilling top drive anomaly suppression method. This method collects data on skipped drilling and stick-slip vibration using sensors, calculates the severity of the condition using a preset formula, and compares it with a preset normal range to determine if an anomaly has occurred. When an anomaly occurs, the top drive motor speed, hook load, or the top drill pipe torque is adjusted using a PID controller. These existing solutions represent typical approaches in the current field of top drive torque control.

[0004] However, the technical solution disclosed in patent CN119900457A directly inputs sensor data into the model without providing any insights into unified semantic identification and high-speed standardized fusion management of multi-source heterogeneous data. When faced with sensors of different protocols and formats, it is difficult to achieve efficient data aggregation and correlation analysis. Its neural network model is used to directly predict torque values ​​instead of jointly modeling multi-scale local time series and global dependencies on long-term series. It cannot output probabilistic probability of anomalies and specific types of anomaly symptoms, and its ability to identify weak early symptoms is limited. Torque adjustment depends on predicted values ​​and fixed protection logic. It does not provide a solution for generating a continuous dynamic torque limit curve in real time by combining the rated parameters of the top drive equipment, thermal limit status, and formation equivalent resistance torque. Nor does it involve a strategy for unified execution of multi-system linkage protection such as torque limit, upper limit of rotational speed, drilling pressure limit, and preventive coordinated triggering of the well control system. The patent with publication number CN120487034A relies on preset calculation formulas and fixed normal ranges for anomaly detection. Data acquisition and processing remain in the traditional local mode, lacking semantic unification and lightweight edge-side model reasoning capabilities. Its control output is conventional PID regulation of motor speed or torque, and it also fails to provide a method for generating dynamic torque limit curves based on anomaly probability and symptom classification. Furthermore, it does not teach the use of a gradual backoff approach to restore control parameters and the addition of confidence markers to ensure transparency and safety redundancy in control decisions when anomalies are resolved. In summary, existing technologies struggle to form a complete control process from multi-source data perception, probabilistic risk identification, dynamic torque limiting and multi-system collaborative protection, to smooth recovery with confidence markers. The early warning and execution stages remain disconnected, and the problems of static torque limits and reliance on manual intervention have not been fundamentally resolved.

[0005] Therefore, there is an urgent need for a control method that integrates unified management of multi-source heterogeneous data, lightweight edge intelligent identification, dynamic torque limiting and well control coordination, and reliable and stable recovery, in order to improve the response timeliness, multi-system protection coordination and automation level of the top drive system under complex working conditions. Summary of the Invention

[0006] This invention aims to address the shortcomings of existing technologies by providing a dynamic torque limiting and early warning linkage edge control method and system for top drive systems, as detailed below: 1) In a first aspect, the present invention provides a dynamic torque limiting and early warning linkage edge control method for top drive systems, the specific technical solution of which is as follows: Real-time acquisition of multi-source sensor data reflecting real-time drilling conditions, and unified semantic point table management to obtain target multi-source sensor data; The trained TCN-Transformer model is used to conduct risk assessment on the time series of target multi-source sensor data to obtain the probability of anomaly occurrence and the type of anomaly symptoms. Based on the probability of anomaly occurrence and the type of anomaly symptoms, combined with the rated performance parameters of the top drive system and the target multi-source sensor data, a dynamic torque limit curve is generated. When the probability of this anomaly exceeds a preset threshold, the current upper limit of torque on the dynamic torque limit curve is lowered, the upper limit of the rotational speed of the top drive system is reduced, and the drill string lifting speed or drilling pressure is limited. At the same time, a coordinated control signal is sent to the well control safety system to trigger the well control related devices to enter the prevention state in advance.

[0007] The beneficial effects of the dynamic torque limiting and early warning linkage edge control method for top drive systems provided by this invention are as follows: By standardizing and fusing multi-source sensor data through unified semantic point table management and a distributed soft bus, target multi-source sensor data is obtained, eliminating data barriers between sensors using different protocols. A trained TCN-Transformer model is used to assess the risk of the time series composed of the target multi-source sensor data, outputting the probability of anomalies and the type of anomaly symptoms. This enables early identification of risks such as stuck drill bits and adhesion at the weak symptom stage, compensating for the shortcomings of traditional fixed threshold judgment and manual monitoring in early detection. Based on the probability of anomalies and the type of anomalies, a dynamic torque limit curve is generated by combining the rated performance parameters of the top drive system and real-time operating information. This allows the upper limit of torque to adaptively adjust with the risk level and formation conditions, overcoming the problems of lag in response and reliance on manual experience in static torque limiting strategies. When the probability of an anomaly exceeds a preset threshold, the upper limit of torque is simultaneously lowered, the upper limit of rotational speed is reduced, the drilling pressure or drill string lifting speed is limited, and a coordinated control signal is sent to the well control safety system to trigger the well control related devices to enter the prevention state in advance. This realizes the parallel linkage of multiple systems of top drive limitation and well control protection, and solves the defect of the separation between early warning signal and control execution.

[0008] Based on the above scheme, the dynamic torque limiting and early warning linkage edge control method for top drive systems of the present invention can be further improved as follows.

[0009] Furthermore, the target multi-source sensing data includes at least one of top drive torque, rotational speed, hook load, drilling pressure, mud pump pressure, and drill string vibration signal; the multi-source sensing data includes raw data frames collected by multiple sensors; real-time acquisition of multi-source sensing data reflecting real-time drilling conditions, and unified semantic point table management to obtain target multi-source sensing data, including: aggregating raw data frames to the edge computing unit via a distributed soft bus; performing protocol type identification and register address space parsing on the raw data frames to identify the protocol type corresponding to the raw data frames; and converting the observations corresponding to the target multi-source sensing data in the raw data frames into engineering data identified by standard semantic objects according to a pre-configured semantic-point mapping file, and using the engineering data as the target multi-source sensing data.

[0010] The beneficial effects of adopting the above-mentioned further solution are as follows: By aggregating raw data frames collected from multiple sensors to the edge computing unit via a distributed soft bus, the problem of scattered and difficult-to-centralize heterogeneous protocol data at the drilling site is solved. Protocol type identification and register address space parsing of the raw data frames automatically identify the communication protocols of different sensors, eliminating the need for manual configuration and reducing the complexity of accessing multiple types of sensors. Based on a pre-configured semantic-point mapping file, observations in different formats are converted into engineering data identified by standard semantic objects, serving as target multi-source sensor data. This allows parameters such as top drive torque, rotational speed, hook load, drilling pressure, mud pump pressure, and drill string vibration signals to be understood and invoked under a unified semantic framework, eliminating data heterogeneity barriers and providing consistent and reliable standardized input for subsequent risk assessment and torque limit calculations.

[0011] Furthermore, a risk assessment of the time series composed of target multi-source sensor data is performed using a trained TCN-Transformer model. This includes: dividing the time series composed of target multi-source sensor data into multiple time windows according to a preset sliding window length; extracting at least one feature from the target multi-source sensor data within each time window, including time-domain statistics, envelope, frequency-domain energy, spectral peak stability, and amplitude variation rate, to obtain a feature tensor; inputting the feature tensor into the trained TCN-Transformer model, where the TCN layer captures local temporal features, and the Transformer layer extracts global dependency features across channels and across windows; and outputting the probability of anomaly occurrence and the type of anomaly symptoms based on the local temporal features and global dependency features.

[0012] The beneficial effects of adopting the above-mentioned further scheme are as follows: By dividing the time series of target multi-source sensor data into multiple time windows according to a preset sliding window length, and extracting various features such as time-domain statistics, envelope, frequency-domain energy, spectral peak stability, and amplitude variation rate to form a feature tensor, the expression dimension of input information is enriched, making subtle changes in operating conditions explicit. The TCN layer in the trained TCN-Transformer model captures local temporal features and can keenly identify short-term abrupt changes in torque and vibration signals. The Transformer layer extracts global dependency features across channels and windows, integrates the long-range collaborative change law between multiple parameters such as top drive torque, hook load, and rotational speed, and finally outputs the probability of anomaly occurrence and the type of anomaly symptoms. This achieves advanced identification and classification of drilling risks at the edge, providing an accurate risk assessment basis for the generation of dynamic torque limit curves.

[0013] Furthermore, based on the probability of anomaly occurrence and the type of anomaly symptoms, combined with the rated performance parameters of the top drive system and the target multi-source sensor data, a dynamic torque limit curve is generated. This includes: obtaining the equipment capability boundary of the top drive system, which is determined by the rated torque, short-term torque, thermal limit value, and speed-torque envelope relationship of the top drive system and the rotating system; based on the speed, drilling pressure, mud pump pressure, well depth, friction, and mud parameters in the target multi-source sensor data, the formation equivalent resistance torque model is corrected online. The formation equivalent resistance torque model is used to estimate the torque required for the drill bit to cut under the current formation conditions; within each control cycle, based on the equipment capability boundary and the corrected formation equivalent resistance torque model, combined with the probability of anomaly occurrence and the type of anomaly symptoms, the upper limit value of torque changing over time is calculated, thereby generating a dynamic torque limit curve, and a preset slope constraint is applied to the rate of change of the dynamic torque limit curve.

[0014] The beneficial effects of adopting the above-mentioned further scheme are: obtaining the equipment capacity boundary determined by rated torque, short-term torque, thermal limit value, and speed-torque envelope relationship, ensuring that the dynamic torque limit curve is always within the safe operating range of the top drive system. Based on speed, drilling pressure, mud pump pressure, well depth, friction, and mud parameters, the formation equivalent resistance torque model is corrected online, making the upper limit of torque conform to the actual formation cutting requirements, avoiding blindly limiting torque and affecting drilling efficiency. Each control cycle calculates the upper limit of torque based on the probability of anomaly occurrence and the type of anomaly symptom, generating a dynamic torque limit curve that changes over time, and applying a preset slope constraint to the rate of change, ensuring a smooth transition of the torque limiting action. This allows for timely tightening of protection when risks increase, while also preventing excessively rapid tightening of the torque limit from causing secondary impact on the drill string.

[0015] 2) In a second aspect, the present invention also provides a dynamic torque limiting and early warning linkage edge control system for top drive systems, the specific technical solution of which is as follows: It includes a data acquisition module, a risk assessment module, a curve generation module, and a linkage control module; The data acquisition module is used to: acquire multi-source sensor data that reflects the real-time working conditions of drilling in real time, and perform unified semantic point table management to obtain target multi-source sensor data; The risk assessment module is used to: use the trained TCN-Transformer model to perform risk assessment on the time series of target multi-source sensor data, and obtain the probability of anomaly occurrence and the type of anomaly symptoms; The curve generation module is used to generate a dynamic torque limit curve based on the probability of anomaly occurrence and the type of anomaly symptoms, combined with the rated performance parameters of the top drive system and the target multi-source sensor data. The linkage control module is used to: lower the current upper limit of torque on the dynamic torque limit curve, reduce the upper limit of rotational speed of the top drive system, and limit the drill string lifting speed or drilling pressure when the probability of the anomaly exceeds the preset threshold. At the same time, it sends a coordinated control signal to the well control safety system to trigger the well control related devices to enter the prevention state in advance.

[0016] 3) In a third aspect, the present invention also provides an electronic device, the electronic device including a processor coupled to a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor, so that the electronic device implements any of the above-mentioned dynamic torque limiting and early warning linkage edge control methods for top drive systems.

[0017] 4) In a fourth aspect, the present invention also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements any of the above-mentioned dynamic torque limiting and early warning linkage edge control methods for top drive systems.

[0018] It should be noted that the beneficial effects of the technical solutions of the second to fourth aspects of the present invention and their corresponding possible implementations can be found in the above description of the technical effects of the first aspect and its corresponding possible implementations, and will not be repeated here. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments of the present invention will be briefly introduced below: Figure 1 This is a flowchart illustrating a dynamic torque limiting and early warning linkage edge control method for a top drive system according to an embodiment of the present invention. Figure 2 This is an architecture diagram corresponding to a dynamic torque limiting and early warning linkage edge control method for a top drive system according to an embodiment of the present invention; Figure 3 This is a flowchart of edge-side anomaly detection and linkage control. Figure 4 This is a schematic diagram of a dynamic torque limiting and early warning linkage edge control system for a top drive system according to an embodiment of the present invention. Detailed Implementation

[0020] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0021] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0022] like Figure 1 As shown in the figure, a dynamic torque limiting and early warning linkage edge control method for a top drive system according to an embodiment of the present invention includes the following steps: S1. Acquire multi-source sensor data in real time to reflect the real-time working conditions of drilling, and perform unified semantic point table management to obtain target multi-source sensor data. The target multi-source sensor data includes at least one of top drive torque, rotary speed, hook load, drilling pressure, mud pump pressure, and drill string vibration signal; the multi-source sensor data includes raw data frames collected by multiple sensors; real-time acquisition of multi-source sensor data reflecting real-time drilling conditions, and unified semantic point table management to obtain the target multi-source sensor data, including: aggregating raw data frames to the edge computing unit via a distributed soft bus; identifying the protocol type and register address space of the raw data frames to identify the corresponding protocol type; converting the observations corresponding to the target multi-source sensor data in the raw data frames into engineering data identified by standard semantic objects according to a pre-configured semantic-point mapping file, and using the engineering data as the target multi-source sensor data. The specific implementation process is as follows: S10. Through a distributed soft bus, raw data frames collected by various sensors deployed at the drilling site are aggregated to the edge computing unit. These sensors directly sense the top drive system and the drilling process, generating raw data frames containing information on at least one physical quantity among top drive torque, rotary speed, hook load, drilling pressure, mud pump pressure, and drill string vibration signals. Since the sensors come from different manufacturers, their output raw data frames follow their respective industrial communication protocols. The distributed soft bus, as a unified data transmission channel, can identify and connect these heterogeneous sensor devices, regardless of their physical interface or data link layer protocol. The distributed soft bus actively acquires or passively receives data streams containing raw data frames from each sensor node and transmits these data streams in real time and completely to the designated data receiving port of the edge computing unit, completing the data aggregation at the physical and data link layers. This step provides a centralized, unified data source containing all raw information for subsequent processing.

[0023] S11. The edge computing unit performs protocol type identification and register address space parsing on the raw data frames converged to the edge computing unit. After receiving the data stream from the distributed soft bus, the edge computing unit first performs a protocol type identification process for each data frame or data channel. By examining features such as the frame header structure, specific identifiers, port numbers, or handshake information of the communication session, the edge computing unit automatically determines whether the data frame follows an industrial communication protocol such as Transmission Control Protocol / Internet Protocol, Message Queuing Telemetry Transmission Protocol, Process Field Network Protocol, Process Fieldbus Protocol, Controller Area Network Protocol, Open Platform Communication Unified Architecture Protocol, Modem Bus Protocol, or Addressable Remote Sensor High-Speed ​​Channel Protocol. After confirming the protocol type, register address space parsing is then performed. Based on the identified specific protocol specifications, the payload portion of the data frame or the agreed register address space is analyzed to locate the starting position, length, and arrangement rules of each observed data within the data frame. This analytical step precisely establishes a preliminary correspondence between the original byte sequence and the physical observations at the drilling site, clarifying which data segments belong to the top drive torque and which belong to the hook load or other measurement parameters, providing a precise addressing and extraction basis for subsequent data conversion work.

[0024] S12. Based on the pre-configured semantic-point mapping file, the observations corresponding to the target multi-source sensor data in the original data frame are converted into engineering data identified by standard semantic objects, and this engineering data is used as the target multi-source sensor data. After completing protocol identification and address resolution, the edge computing unit performs data conversion according to the previously configured and distributed semantic-point mapping file. This file contains complete mapping rules from raw data points of specific protocols and addresses to standard semantic objects. The edge computing unit searches for matching records in the mapping file based on the parsed protocol type and register address. Once a matching record is found, the edge computing unit extracts the raw value in binary or other formats from the specified position of the original data frame, and converts this raw value into an engineering value with actual physical meaning according to the preset scaling factor, end-order processing method, data type conversion rules, and engineering units in the mapping rules. For example, an integer value stored in a register may be multiplied by a scaling factor and appended with the unit "kilonewton-meters" to finally generate engineering data representing the standard semantic object "top drive torque". Simultaneously, some quality information carried in the original data frame is mapped to the quality factor of the standard semantic object. Through this transformation step, the multi-source heterogeneous raw sensor data is thoroughly cleaned, standardized, and uniformly presented in the form of standard semantic objects. The resulting set of engineering data, identified and organized using standard semantic objects, constitutes the acquired target multi-source sensor data. It shields the complexity of the underlying sensors and protocols, providing high-quality, semantically consistent standardized input for subsequent time-series analysis and risk assessment.

[0025] The distributed soft bus is a data communication infrastructure built at the operating system level, logically connecting physically dispersed devices into a unified communication network. In this technical solution, the distributed soft bus shields the differences in underlying physical connection methods and communication protocols, providing a high-speed, reliable data channel for various sensors and edge computing units deployed at the drilling site. It enables sensor devices from different manufacturers and with different interface types to be discovered, connected, and exchange data in real time, achieving transparent transmission and aggregation of multi-source sensor data, and serving as the physical foundation for the entire data flow.

[0026] Unified semantic point table management is a data standardization and abstraction management method. It establishes a globally consistent description standard, defining a unique and unambiguous standard semantic object for each physical quantity or equipment status at the drilling site. Through this management method, regardless of the communication protocol or data format used by the underlying sensors, their physical meanings, such as top drive torque or hook load, will be mapped and uniformly identified as the same standard semantic object. This allows the upper-level data processing model and control logic to focus on standardized engineering data for analysis and decision-making without needing to concern themselves with the data source or original format, thus eliminating the semantic gap between multi-source heterogeneous data.

[0027] The edge computing unit is a computing device deployed at the drilling site, close to the data source. It has the capability to process data and run intelligent models, responsible for real-time analysis and decision-making on the multi-source sensor data aggregated in this technical solution. The edge computing unit does not rely on a remote cloud data center and can complete tasks such as data preprocessing, feature extraction, and deep learning model inference locally, thus ensuring a millisecond-level rapid response to real-time operating conditions even under network latency or interruptions that may occur at the drilling site.

[0028] Protocol type identification refers to the process of automatically determining the industrial communication protocol standard followed by the raw data frames received from different sensors. Due to the wide variety of sensors in drilling sites, they may employ one of several protocols, such as Transmission Control Protocol / Internet Protocol, Message Queuing Telemetry Transmission Protocol, Process Field Network Protocol, Process Fieldbus Protocol, Controller Area Network Protocol, Open Platform Communication Unified Architecture Protocol, Modem Bus Protocol, or Addressable Remote Sensor High-Speed ​​Channel Protocol. Accurately identifying the protocol type corresponding to each data frame is a prerequisite for subsequent correct data parsing and information extraction.

[0029] Register address space resolution is a specific data processing step following protocol type identification. After determining the communication protocol used by an original data frame, it is necessary to analyze and locate the register addresses or node address spaces within the data frame according to the protocol's specifications. This process clarifies which byte or specific register address in the data frame corresponds to a particular physical observation, such as the top drive torque or hook load, thus providing precise guidance for subsequently extracting and converting this value into physically meaningful engineering data.

[0030] The semantic-point mapping file is a pre-configured, structured set of description rules. It details the correspondence between different industrial protocol types, different register addresses, and standard semantic objects defined in the unified semantic point table. This file defines how to convert and map raw data points with protocol characteristics to a standard semantic object with clear engineering meaning, and includes attribute information such as data type, scaling rules, engineering units, and read / write permissions. This file is the key basis for achieving automated conversion from multi-source heterogeneous raw data to standardized engineering data.

[0031] A standard semantic object is an abstract data unit defined in the unified semantic point table management system. It describes a measurement or control variable with clear engineering meaning at the drilling site. Each standard semantic object has a complete set of attribute definitions, including a unique name identifier, data type, unit, measurement range, and quality identifier. For example, top drive torque would be defined as a standard semantic object named "Top Drive Torque". All internal system access and processing of top drive torque data is based on this standard semantic object, ensuring data consistency and unambiguity throughout the entire process.

[0032] S2. Utilize the trained TCN-Transformer model to perform risk assessment on the time series data composed of multi-source sensor data of the target, obtaining the probability of anomaly occurrence and the types of anomaly symptoms, specifically including: The time series of target multi-source sensor data is divided into multiple time windows according to a preset sliding window length. For each time window, at least one feature is extracted from the target multi-source sensor data, including time-domain statistics, envelope, frequency-domain energy, spectral peak stability, and amplitude variation rate, to obtain a feature tensor. This feature tensor is then input into a trained TCN-Transformer model. The TCN layer in the trained TCN-Transformer model captures local temporal features, and the Transformer layer extracts global dependency features across channels and windows. Based on the local temporal features and global dependency features, the probability of anomaly occurrence and the type of anomaly symptom are output. The specific implementation process is as follows: S20. Divide the time series composed of target multi-source sensor data into multiple time windows according to a preset sliding window length. After transmitting the target multi-source sensor data to the edge computing unit and completing the standardization, these data are arranged into a continuous multivariate time series composed of various physical quantities according to a unified timestamp. Each column in this multivariate time series corresponds to one type of target multi-source sensor data, such as one or more columns from the top drive torque column, rotational speed column, hook load column, drill pressure column, mud pump pressure column, and drill string vibration signal column; each row corresponds to a sampling time. When dividing the time window, a predetermined preset sliding window length, such as 128 data points, is used as the window size, and the data is truncated along the time axis from the past to the present with a fixed sliding step size. The sliding step size is usually set to 1 data point or a small value to ensure that adjacent time windows have a large overlap area, thereby achieving quasi-continuous extrapolation of risk identification. With each sliding step, a total of preset sliding window length of continuous sampling points are truncated from the current time backward to form a time window. This time window contains a two-dimensional data matrix, with the number of rows equal to the preset sliding window length and the number of columns equal to the number of channels in the target multi-source sensor data. Multiple time windows are arranged in chronological order to form a processing sequence, which is then used sequentially by subsequent feature extraction steps.

[0033] S21. Extract at least one feature from the target multi-source sensing data within each time window, including time-domain statistics, envelope, frequency-domain energy, spectral peak stability, and amplitude variation rate, to obtain a feature tensor. For a given two-dimensional data matrix within a time window, the feature extraction process is performed independently along each target multi-source sensing data channel, and the cross-relationships of multiple channels are also statistically analyzed simultaneously, specifically including some or all of the following five features.

[0034] The extraction method for temporal statistical features is as follows: For a time series of a single channel, all values ​​within a preset sliding window length are traversed, and the maximum, minimum, arithmetic mean, variance, standard deviation, and root mean square value are calculated cumulatively. Simultaneously, skewness is calculated to measure the degree to which the numerical distribution deviates from symmetry, and kurtosis is calculated to measure the degree to which the numerical distribution is flat or sharp. For the target multi-source sensor data of twelve channels, a set of temporal statistical features is generated for each channel.

[0035] The envelope feature extraction method is as follows: For a time series of a channel, an imaginary part sequence orthogonal to the original sequence is first constructed using Hilbert transform. The original sequence is used as the real part. The square root of the sum of the squares of the real and imaginary parts is then taken to obtain the amplitude sequence of the analytic signal; this amplitude sequence is the upper envelope. After the envelope sequence is generated, its mean, standard deviation, and fluctuation frequency are calculated as aggregated features, which are used as the envelope features of that channel. When computational resources are limited, a peak-preservation and low-pass filtering cascade method can be used to approximate the signal's outer contour instead of analytical envelope extraction.

[0036] The extraction method for frequency domain energy features is as follows: For a time series of a channel, a Fast Fourier Transform is performed to obtain the complex spectrum of the time series in the frequency domain. The amplitude spectrum is obtained by taking the modulus of the complex spectrum. The frequency axis corresponding to the amplitude spectrum is divided into three intervals: low frequency, mid frequency, and high frequency. The sum of the squares of the amplitude spectrum values ​​in different frequency bands is calculated to obtain the frequency domain energy value of that band. Alternatively, the energy of each frequency band can be divided by the total energy of the entire frequency band to obtain the normalized frequency band energy ratio. For channels with rich frequency domain information, such as drill string vibration signals, frequency domain energy features are particularly important.

[0037] The extraction method for spectral peak stability features is as follows: The spectrum of the same channel in the previous time window is compared. The frequency positions corresponding to several maximum peaks in the current time window spectrum are recorded, forming a spectral peak frequency set. The spectral peak frequency set of the current time window is matched with the spectral peak frequency set of the previous time window to find corresponding peak pairs with frequency position deviations within the allowable range. For each matched peak pair, the frequency drift and amplitude change rate are calculated, and the root mean square of the amplitude change rates of all matched peaks is used as the spectral peak stability feature. If a new spectral peak appears in the current time window spectrum that did not appear in the previous time window and whose amplitude exceeds a predetermined significant threshold, it is also considered a degradation event of spectral peak stability, and a discrete marker feature representing spectral peak instability is directly generated.

[0038] The amplitude variation rate feature is extracted as follows: For a time series of a channel, the first-order difference absolute value sequence between adjacent sampling points in the sequence is calculated. The mean, standard deviation, and percentage of counts exceeding a preset jump threshold are then calculated from this first-order difference absolute value sequence. The mean reflects the average amplitude change of the signal, the standard deviation reflects the dispersion of the amplitude change, and the percentage of counts exceeding the preset jump threshold directly characterizes the frequency of sudden, large jumps in the signal.

[0039] Within a time window, the various features extracted for each channel are concatenated into a long one-dimensional feature vector in a predetermined order. Since each channel extracts several items from the five feature categories mentioned above, and the number of channels is fixed, the one-dimensional feature vectors of all channels are then stacked along the channel dimension to form a two-dimensional feature tensor with a shape equal to the number of channels multiplied by the number of features per channel. This feature tensor serves as a comprehensive representation of all features of the target multi-source sensing data within that time window, and is also the standard input format for the trained TCN-Transformer model.

[0040] S22. The feature tensor is input into the trained TCN-Transformer model, where the TCN layer captures local temporal features. After being fed into the model, the feature tensor first enters the TCN layer, which is also the temporal convolutional network layer. The TCN layer consists of several stacked residual blocks. In the first residual block, the feature tensor undergoes a first-level dilated causal convolution operation along the time dimension. The convolution kernel slides unidirectionally along the time axis, and each output time step only contains information from the current and previous input time steps, ensuring causality. The dilation coefficient causes the receptive field of the convolution kernel to expand exponentially. Shallow residual blocks capture extremely short-term local features within a range of several adjacent data points, while deep residual blocks capture longer-term local features spanning tens or even hundreds of data points. After each convolution operation, the convolution kernel parameters are reparameterized through weight normalization to stabilize training and inference. A nonlinear transformation is introduced through an activation function layer, and some intermediate neurons are randomly discarded through a dropout regularization layer to enhance generalization ability. The identity mapping path within the residual block directly or through 1×1 convolution adjusts the dimension of the input and then superimposes it onto the output of the main transform path before activation, completing one residual learning cycle. After layer-by-layer processing of all residual blocks, the TCN layer outputs a local temporal feature sequence that inherits the temporal dimension of the original feature tensor but deeply incorporates multi-scale local surrounding information at each time step. In this local temporal feature sequence, local abrupt change patterns such as short-duration torque pulses and instantaneous vibration impacts in the drilling data are effectively captured and encoded.

[0041] S23. Global dependency features across channels and windows are extracted from the Transformer layer of the trained TCN-Transformer model. The local temporal feature sequence is then output and fed into the Transformer layer. At the front end of the Transformer layer, positional encoding is superimposed on each time step vector of the local temporal feature sequence. This is done by adding a learnable positional vector corresponding to the time step position and matching its dimension element-wise to the original feature vector, giving the model the ability to recognize the sequence order. The sequence after superimposed positional encoding passes through multiple Transformer encoder layers sequentially. In each Transformer encoder layer, a multi-head self-attention sublayer first generates a set of query vectors, a set of key vectors, and a set of value vectors for each time step vector in the sequence through different linear mappings. Self-attention calculation matches the query vectors with the key vectors of all time steps to derive the relevance weight of each time step to all time steps in the entire sequence. This weight is applied to the value vector, completing the fully connected modeling of the relationship between any two time steps in the input sequence. Since the generation of query, key, and value vectors corresponds to the linear combination of different sensor parameter channel features, the self-attention mechanism can automatically discover the interaction relationships between different target multi-source sensor data channels while calculating global temporal dependencies. For example, the correlation between abnormal changes in top drive torque and the synchronous rise of hook load is captured by high attention weights and encoded into the output vector. The multi-head attention mechanism sets up multiple sets of parallel query, key, and value transformations. Each set is calculated independently, concatenated, and then fused through linear transformation to obtain an enhanced representation containing multi-angle cross-channel and cross-window dependency information. The output after self-attention calculation is added to the input through residual connections and subjected to layer normalization. Then, it enters a feedforward fully connected sub-layer consisting of two linear layers and an activation function layer for further transformation. Finally, it completes the entire processing of an encoder layer through residual connections and layer normalization. After being processed by multiple encoder layers, the representation vector of each time step in the sequence not only contains its own local temporal characteristics, but also internalizes the complex interaction relationships of all positions and all sensor data channels within the entire time window and even beyond the range of a single time window. These representation vectors are the finally extracted global dependency features across channels and across windows.

[0042] S24. Output the probability of anomaly occurrence and the type of anomaly symptom based on local temporal features and global dependency features. After processing by the Transformer layer, the global representation vector of the sequence, obtained by average pooling compression of the representation vector corresponding to the last time step in the sequence or all time step representation vectors, comprehensively carries the multi-scale local temporal features extracted by the TCN layer and the cross-channel and cross-window global dependency features refined by the Transformer layer. This global representation vector of the sequence is passed to the output segment of the model. The output segment has two parallel output paths. In the first path, the global representation vector of the sequence is mapped to a single real value through a fully connected linear layer, and then mapped to the range of 0 and 1 through a sigmoid activation function. The output is the probability of anomaly occurrence, representing the risk of anomalies occurring in the drilling process within the time period covered by the current time window. In the second path, the global representation vector of the sequence is mapped to a vector with a dimension equal to the number of predefined anomaly symptom categories through another fully connected linear layer. This vector is converted into a category probability distribution through a Softmax normalization function. The category label corresponding to the component with the largest value in the distribution is the type of anomaly symptom, such as "pre-sticking" or "adhesion". The probability of anomaly occurrence and the type of anomaly symptoms output based on local time-series features and global dependency features are used as risk assessment results for direct use in subsequent dynamic torque limiting calculations and linkage control steps.

[0043] The preset sliding window length is a pre-defined integer value used to define the number of consecutive data points contained within a time window. This value determines the length of the historical time series interval observed each time the trained TCN-Transformer model is input for risk assessment. In drilling applications, the preset sliding window length needs to balance sensitivity to rapid, sudden anomalies with coverage of slow, evolving trends. Typical values ​​are set to 128 or 256 data points, corresponding to a data sampling frequency of 256 Hz to 1000 Hz, with a corresponding time span of approximately a fraction of a second to a few seconds.

[0044] Time-domain statistics are a method for calculating the numerical distribution characteristics of multi-source sensor data of a target within a time window, directly on the time axis. For multi-source sensor data of a specific target within the window, such as top drive torque or hook load, statistical measures such as mean, variance, standard deviation, peak value, root mean square value, skewness, and kurtosis are calculated to reflect the average level, fluctuation intensity, extreme value magnitude, and symmetry and steepness of the distribution pattern of the sensor data within the window period. When drilling conditions become abnormal, these time-domain statistics often undergo identifiable changes; for example, the variance of torque may increase significantly, and the root mean square value may rise simultaneously.

[0045] The envelope is a feature signal extracted from the upper and lower contours of the target multi-source sensor data time series. For data with high-frequency fluctuation characteristics, such as top drive torque or drill string vibration signals, the analytic signal can be obtained through Hilbert transform or its simplified implementation, and then the amplitude of the analytic signal can be obtained to obtain the upper envelope. Alternatively, the envelope can be extracted through peak detection and interpolation smoothing. The envelope feature can characterize the overall shape and low-frequency modulation pattern of the fluctuation amplitude of the target multi-source sensor data, which helps to detect slow periodic fluctuations or trend changes hidden in the original high-frequency signal. For example, when the friction state between the drill string and the formation changes, the envelope of the torque signal may show regular fluctuations.

[0046] Frequency domain energy is an energy distribution index calculated across different frequency ranges after the time series of multi-source sensor data of the target within a time window is transformed to the frequency domain using a Fast Fourier Transform. Frequency domain energy is not a single numerical value, but rather a set of characteristics, corresponding to the energy proportion or absolute energy value in the low-frequency, mid-frequency, and high-frequency bands, respectively. When specific anomalies occur in the drill string, the frequency domain energy distribution of signals such as vibration will undergo characteristic shifts. For example, mechanical imbalance may lead to a significant concentration of energy at the rotational frequency, while stick-slip phenomena may generate abnormal energy clusters in lower frequency bands.

[0047] Among them, spectral peak stability is a characteristic that measures the stability of the position and amplitude of the main spectral peaks on the frequency axis over time in the spectrum. By tracking and comparing the spectra of multiple consecutive sub-windows or adjacent time windows, the frequency drift and amplitude fluctuation of each main spectral peak are recorded, and a stability index is calculated comprehensively. During normal and stable drilling, the fundamental frequency of the rotational speed and its harmonics corresponding to the main spectral peaks usually remain relatively stable; however, when signs of stuck pipe or well kick occur, the spectral peaks may show frequency shifts, violent amplitude jumps, or even the sudden appearance of new unstable spectral peaks, and the spectral peak stability will decrease accordingly.

[0048] Amplitude variability is a characteristic describing the degree of variation and the intensity of instantaneous changes in the amplitude of multi-source sensor data within a time window. This characteristic is obtained by calculating the statistical distribution of amplitude differences between adjacent sampling points within the time window, the frequency of sign changes, or the cumulative value of absolute changes. When the drilling process is in a stable state, the amplitude changes of adjacent torque or vibration signals are relatively gentle; when abnormal friction, irregular collisions of the drill string, or sudden changes in the formation occur, the amplitude variability will increase significantly, showing frequent and large-amplitude rapid jumps in the signal.

[0049] The feature tensor is a multi-dimensional data structure formed by organizing and arranging all features extracted within a time window. In this feature tensor, different dimensions correspond to different time steps within the time window, different target multi-source sensor data channels, and different types of features, such as time-domain statistical features, envelope features, frequency-domain energy features, spectral peak stability features, and amplitude variation rate features. The trained TCN-Transformer model receives this structured feature tensor as input, enabling it to treat multi-source, multi-class features as a unified whole for subsequent pattern recognition and risk inference.

[0050] Among them, the global dependency features across channels and across windows are a type of high-level abstract feature extracted by the Transformer layer in the trained TCN-Transformer model. Cross-channel refers to the temporally synchronous correlation between multi-source sensor data from different targets, such as the correlation between top drive torque and hook load, or between top drive torque and drill string vibration signal. Cross-window refers to the long-term dependency between multiple time windows that are far apart in time dimension; for example, there may be a causal relationship between a slight torque fluctuation that occurred a few minutes ago and a sudden stick-slip trend at the present moment. Global dependency features encode the complex interaction information of the above cross-channel and cross-window features into a set of high-dimensional vector representations, overcoming the limitation of local temporal features that only focus on data patterns within short time windows.

[0051] The specific implementation process for obtaining the trained TCN-Transformer model is as follows: S020. Construct the network layer structure of the TCN-Transformer model. As a hybrid deep learning architecture, the network layer structure of this model consists of three interconnected parts: the backbone extraction section, the global modeling section, and the output section. Each part is composed of several functional network layers stacked together.

[0052] The backbone extraction segment consists of one or more stacked temporal convolutional network residual blocks. Each temporal convolutional network residual block contains two branches: a main transformation path and an identity mapping path. In the main transformation path, data flows sequentially through a first dilated causal convolutional layer, a first weight normalization layer, a first activation function layer, a first dropout regularization layer, a second dilated causal convolutional layer, a second weight normalization layer, a second activation function layer, and a second dropout regularization layer. The kernel size of each dilated causal convolutional layer can be set to a small odd number, such as 3 or 5, depending on the temporal granularity of the drilling data. The dilation coefficient increases exponentially between different residual blocks, starting from 1 and increasing to 2, 4, 8, to progressively expand the temporal sensing range layer by layer. This allows shallow layers to capture sudden torque or vibration changes within very short intervals, while deeper layers cover slow parameter drift over longer time ranges. The padding method of the causal convolutions ensures that the output of each time step is only related to the current location and its historical data. The identity mapping path is a 1×1 convolutional layer. When the input and output dimensions of the main transformation path do not match, this 1×1 convolutional layer adjusts the input dimension to match the output; otherwise, it directly performs element-wise addition. The residual connection adds the output of the main transformation path to the output of the identity mapping path. The result is passed through a final activation function layer and serves as the complete output of the residual block of this temporal convolutional network. Multiple such residual blocks are stacked sequentially to form a complete backbone extraction segment, used to extract high-level feature sequences that incorporate multi-scale local temporal texture information from the input feature tensor layer by layer.

[0053] The global modeling segment, following the output of the backbone extraction segment, consists of a positional encoding layer and multiple cascaded Transformer encoder layers. The positional encoding layer superimposes a unique and learnable positional vector onto each time step vector in the high-level feature sequence output from the backbone extraction segment. This injects absolute or relative positional information into the sequence, compensating for the inherent limitation of the Transformer structure itself in recognizing temporal order. The sequence with superimposed positional encoding then enters the first Transformer encoder layer. The processing flow within a Transformer encoder layer is as follows: the input sequence first enters a multi-head self-attention sub-layer. In this sub-layer, the input sequence is mapped into query vectors, key vectors, and value vectors through different linear transformation matrices. The dot product similarity between the query vector and the key vector is then calculated and normalized to attention weights using the Softmax function. These weights are then used to perform a weighted summation of the value vectors, thereby generating a representation vector for each time step in the sequence that incorporates global contextual information. The multi-head self-attention mechanism employs multiple sets of parallel linear transformation matrices for query vectors, key vectors, and value vectors. Each set independently performs the attention calculations, ultimately concatenating the outputs of all attention heads and mapping them back to the original dimension through a linear transformation to fuse global dependencies across different representation subspaces. The output of the multi-head self-attention sublayer is added to the input via a residual connection and then normalized. The normalized result enters a feedforward fully connected sublayer, which consists of two linear transformation layers and an activation function layer, performing independent nonlinear transformations on the representation of each position in the sequence. The output of the feedforward fully connected sublayer is also added to the input via a residual connection and normalized again, forming the final output of the Transformer encoder layer. Multiple Transformer encoder layers with identical structures are sequentially connected, enabling multiple combinations of the self-attention mechanism at different levels of abstraction and repeated refinement of deep-level global dependencies.

[0054] The output segment, connected after the last Transformer encoder layer, generates two outputs: the probability of anomaly occurrence and the type of anomaly symptom. Internally, the output segment contains a pooling path for global sequence representation and two bifurcated parallel linear layers. The pooling path extracts the representation vector from the last time step of the sequence (corresponding to the output of a Transformer encoder layer) or performs average pooling on the representation vectors of all time steps in the entire sequence, compressing them into a fixed-dimensional global sequence representation vector. This global sequence representation vector is simultaneously input to two parallel linear layers. One layer is connected to a sigmoid activation function, outputting a scalar between 0 and 1, representing the probability of anomaly occurrence. The other layer is connected to a softmax normalization function, outputting a probability distribution vector with a length equal to the total number of predefined anomaly symptom categories. Each element in the vector represents the probability of the corresponding anomaly symptom type; typically, the category with the highest probability is selected as the final anomaly symptom type. At this point, the complete network layer structure of the TCN-Transformer model is constructed. This structure has the ability to extract local temporal texture from the original input features, model global dependencies, and simultaneously output quantitative and qualitative risk categories.

[0055] S021. Train the completed TCN-Transformer model to obtain the trained model.

[0056] The training process begins with acquiring labeled historical drilling data. Historical multi-source sensor data is organized into standardized target multi-source sensor data time series using the same unified semantic point table management method as described above. Drilling engineering experts manually label specific time windows within these time series based on historical operation records, accident reports, and experience. Each time window is assigned a binary label indicating the presence of an actual anomaly within that window, and an anomaly symptom type label, such as whether a stuck pipe or adhesion event occurred within the corresponding window's time period. The labeled data is then organized into continuous sliding window samples in chronological order. Each sample contains a feature tensor composed of target multi-source sensor data within a time window, along with corresponding anomaly presence / absence labels and anomaly symptom type labels, forming the sample dataset required for supervised training.

[0057] Before starting model training, the training environment must be determined to be either an edge computing unit simulation environment or a general-purpose graphics processing unit server, and training hyperparameters must be configured, including batch size, initial learning rate, learning rate decay strategy, maximum number of training epochs, and early stopping patience value. All learnable parameters of the model, including the convolution kernels and biases of each convolutional layer, the weight matrices and bias vectors of each linear layer, the scaling and translation parameters of each normalized layer, the vector table of the positional encoding layer, and all linear transformation matrices in multi-head self-attention, are initialized according to a unified random initialization strategy.

[0058] The training process for each round is as follows: Training samples are divided into several batches. Samples within each batch are read in and undergo forward propagation processing identical to that used in actual model inference. This involves layer-by-layer local feature extraction from the residual blocks of the temporal convolutional network, positional encoding stacking, multi-head self-attention and feedforward fully connected transformations in the Transformer encoder layer, and finally, two linear layers—sequence global representation extraction and output segment—to obtain the predicted probability of anomalies and the predicted distribution of anomaly symptom types for each sample within each batch. When calculating the loss, a multi-task joint loss function is used, consisting of a weighted sum of two parts. One part is a binary cross-entropy loss for the probability of anomalies, comparing the error between the predicted probability value of each sample and the actual presence or absence of anomaly labels. The other part is a multi-class cross-entropy loss for the anomaly symptom types, comparing the error between the predicted probability distribution of each sample and the actual anomaly symptom type labels. The weighting coefficients of the two loss parts are pre-adjusted based on model validation performance to ensure effective optimization of both types of outputs during training.

[0059] After the loss value is calculated, it is propagated forward layer by layer from the output layer using the backpropagation algorithm, calculating the gradient of the loss value with respect to all learnable parameters in each layer. Based on the selected stochastic gradient descent or its variant optimizer, such as the adaptive moment estimator, the parameter values ​​are iteratively updated using the gradient values ​​of each parameter and the current learning rate. Each batch completes one parameter update, and a training epoch is completed when all batches have been traversed once.

[0060] After each training epoch, the current model is validated using an independently partitioned validation set. The data format of the validation set is the same as that of the training set, but the drilling intervals and time ranges it covers do not overlap with the training samples. Forward propagation is also performed on each sample in the validation set to obtain the predicted probability of anomaly occurrence and the type of anomaly symptom, and the multi-task joint loss value on the validation set, as well as additional evaluation metrics such as precision, recall, and F1 score, are calculated. When the validation set loss value no longer decreases or even increases after several consecutive training epochs, an early stopping mechanism is triggered, the training process terminates, and the model parameters corresponding to the epoch with the smallest validation set loss value are selected as the final saved model parameters.

[0061] After the training is completed, the model structure definition file and the finally saved model parameter file are stored in a serialized manner together, and can be further converted into a lightweight inference format compatible with the hardware platform of the edge computing unit through a model compression or quantization tool. Through the entire training process described above, the obtained model with a complete network structure and an optimized parameter set is the trained TCN-Transformer model, which can be directly used for risk assessment of real-time drilling data, and output the occurrence probability of anomalies and the types of anomaly signs.

[0062] Wherein, the TCN-Transformer model is a deep learning hybrid architecture for processing time series data. The two network structures of the temporal convolutional network and the Transformer are organically integrated. The temporal convolutional network is good at capturing short-term fluctuations and fine-grained texture features from the local window of the time series, while the Transformer can model the long-distance dependency relationship between data points far apart in the sequence by virtue of the self-attention mechanism. In drilling condition analysis, the instantaneous jitter of torque data and the slow collaborative change trend among multiple parameters such as hook load and rotational speed need to be identified by the above two capabilities respectively. Therefore, the hybrid architecture can achieve high-precision identification of complex drilling anomaly signs under the limited computing resources of the edge computing unit.

[0063] Wherein, the temporal convolutional network layer is a convolutional neural network structure specially designed for sequence modeling. This network layer adopts a causal convolution method to ensure that the output at time t only depends on the input data at time t and previous moments, and does not use future information, which ensures the correctness of the causal relationship of real-time data processing during the risk assessment process. At the same time, the temporal convolutional network layer introduces dilated convolution technology, which increases the receptive field through interval sampling, so that the network can cover a longer time history interval without adding too much computation. Residual connections are also embedded in each temporal convolutional network block to alleviate the gradient disappearance problem during deep network training and ensure the stability of model training.

[0064] Wherein, the anomaly occurrence probability is a continuous value between 0 and 1 output by the TCN-Transformer model, which is used to quantify the possibility that a certain abnormal working condition is occurring or is about to occur in the drilling process within the current sliding time window. The closer the value is to 1, the more confident the model is in judging that there is a current risk; the closer the value is to 0, the more the model considers the working condition to be close to a normal state. This probability value is the direct trigger basis for subsequent dynamic torque limit calculation and linkage control decision-making.

[0065] The abnormal sign type is a classification label output by the TCN-Transformer model, indicating the specific risk category identified by the model. In top drive drilling scenarios, possible abnormal sign types include labels such as "precursor to stuck pipe," "signs of sticking," "slippage tendency," and "mechanical imbalance." This information, used in conjunction with the probability of anomaly occurrence, can guide the selection of subsequent dynamic torque limiting and linkage control strategies, enabling differentiated control measures to be taken for different types of risks.

[0066] S3. Based on the probability of anomaly occurrence and the type of anomaly symptoms, combined with the rated performance parameters of the top drive system and the target multi-source sensor data, a dynamic torque limit curve is generated, specifically including: The equipment capability boundary of the top drive system is obtained. This boundary is determined by the rated torque, short-term torque, thermal limit value, and speed-torque envelope relationship of the top drive system and the rotating system. Based on the target multi-source sensor data (speed, drilling pressure, mud pump pressure, well depth, friction, and mud parameters), the formation equivalent resistance torque model is corrected online. This model is used to estimate the torque required for the drill bit to cut under current formation conditions. Within each control cycle, based on the equipment capability boundary and the corrected formation equivalent resistance torque model, combined with the probability of anomaly occurrence and the type of anomaly symptoms, the upper limit value of torque changing over time is calculated, thereby generating a dynamic torque limit curve. A preset slope constraint is applied to the rate of change of the dynamic torque limit curve. The specific implementation process is as follows: S30. Obtain the equipment capability boundary of the top drive system. The equipment capability boundary is determined by the rated torque, short-time torque, thermal limit value, and speed-torque envelope relationship of the top drive system and the rotating system.

[0067] The rated torque is obtained by directly reading it from the equipment technical manual, nameplate parameters, or electronic parameter table stored in the control equipment provided by the top drive system manufacturer. This rated torque value remains unchanged under standard operating conditions and serves as the basic reference line for the equipment's capability boundary.

[0068] The short-term torque is obtained by extracting short-term overload capacity parameters from the top drive system's technical specifications, acquiring the numerical multiple of the short-term torque and its corresponding allowable duration. For example, if the extracted short-term torque is 150% of the rated torque, it can be allowed to last for 60 seconds. These two parameters are linked and recorded for subsequent dynamic calculations to determine how long the current torque level has exceeded the rated torque. Once the cumulative duration exceeds the allowable short-term torque, the equipment's capacity boundary will be forcibly tightened to near the rated torque.

[0069] The thermal limit value is obtained by reading real-time data from temperature sensors installed in key heat-generating components such as the top drive motor windings, gearbox lubrication lines, and main bearings, or by using a thermodynamic temperature rise model established based on historical load current, ambient temperature, and heat dissipation conditions to estimate the current temperature of the heat-generating components. This estimated temperature is compared with the manufacturer-specified maximum allowable operating temperature for each component. If the current temperature is below the alarm threshold, the thermal limit value is set at the rated torque or short-term torque. If the temperature approaches or exceeds the alarm threshold, the thermal limit value is gradually reduced according to a preset derating curve; the higher the temperature, the lower the allowable torque limit. The thermal limit value is recalculated in each control cycle to reflect the equipment's thermal constraints in real time.

[0070] The speed-torque envelope relationship is obtained by reading the drive characteristic curves or digital models provided by the top drive system supplier. This relationship is typically stored as a two-dimensional table or a piecewise linear function, where the horizontal axis represents the rotational speed and the vertical axis represents the maximum allowable continuous output torque at that speed. During actual operation, based on the current rotational speed value from the target multi-source sensor data, the upper limit of the envelope torque at the current speed is obtained by looking up the value in the two-dimensional table or through linear interpolation. If the top drive system uses a variable frequency motor, the speed-torque envelope relationship also includes the boundary speed between the constant torque region and the constant power region, as well as the corresponding torque inflection point.

[0071] The rated torque, short-time torque, thermal limit value, and the upper limit of the envelope torque corresponding to the current speed are collected, and the minimum value among the four is taken in each control cycle as the absolute upper limit of torque allowed by the equipment capacity boundary at the current moment. The generation of subsequent dynamic torque limit curves must not exceed this upper limit value.

[0072] S31. Based on the target multi-source sensor data, including rotational speed, drilling pressure, mud pump pressure, well depth, friction, and mud parameters, the formation equivalent resistance torque model is corrected online. The formation equivalent resistance torque model is used to estimate the torque required for the drill bit to cut under the current formation conditions.

[0073] The formation equivalent resistance torque model is established based on an initial foundation model. This foundation model is pre-constructed based on drilling engineering design data. The input parameters include static or quasi-static parameters such as drill bit diameter, drill bit type coefficient, rock drillability grade of the drilled formation, design well depth, and mud density design value. The output is the estimated torque required for normal drilling under design conditions.

[0074] During real-time drilling, the target multi-source sensor data, including rotational speed, drilling pressure, mud pump pressure, well depth, friction, and mud parameters, are continuously collected and used to implement online corrections to the basic model. The correction methods are as follows: The measured drill pressure is compared with the model's estimated drill pressure. If the measured drill pressure is higher than the designed drill pressure, it indicates that the depth of the drill bit's cutting teeth into the rock has increased, and the required cutting torque needs to be increased. The adjustment amount is proportional to the drill pressure deviation. The measured mud pump pressure is compared with the benchmark pump pressure under normal circulation conditions. An increase in pump pressure usually means increased cuttings resistance at the bottom of the well or poor annulus flow, and the equivalent resistance torque is adjusted accordingly. Well depth parameters are obtained from real-time well depth data obtained from the drilling rig control system. As the well depth continues to increase, the contact area and friction between the drill string and the wellbore increase accordingly. The torque component corresponding to the frictional work in the model is accumulated online using an empirical coefficient per meter increment. Frictional parameters can be indirectly estimated based on the difference between the rotary table torque and the drill bit torque. The frictional factor is extracted and updated using torque fluctuation data during each single-joint or reaming process. Mud parameters include mud density, viscosity, and solid content. When mud parameters change, the buoyancy of the drill string, flow resistance, and cuttings carrying efficiency are all affected, and the corresponding correction coefficients in the model are updated synchronously.

[0075] The deviations or real-time values ​​of the above parameters are summarized according to pre-calibrated weights and transfer functions to generate a correction coefficient. This correction coefficient is then multiplied by the estimated base torque output from the basic model to obtain the online corrected formation equivalent resistance torque. This correction is performed once in each control cycle to ensure that the formation equivalent resistance torque model remains consistent with the current actual formation conditions and drilling parameters.

[0076] S32. In each control cycle, based on the equipment capacity boundary and the modified formation equivalent resistance torque model, combined with the probability of anomaly occurrence and the type of anomaly symptoms, calculate the upper limit value of torque that changes over time, thereby generating a dynamic torque limit curve, and apply a preset slope constraint to the rate of change of the dynamic torque limit curve.

[0077] Within each control cycle, the acquired equipment capability boundary is first read to obtain an absolute torque upper limit that will not be exceeded. At the same time, the predicted resistance torque output by the corrected formation equivalent resistance torque model at the current moment is read. This predicted value is the basic torque value required by the drill bit under the current normal cutting conditions.

[0078] The dynamic safety margin is calculated by combining the probability of anomaly occurrence and the type of anomaly symptoms. The probability of anomaly occurrence is a continuous value between 0 and 1. Multiplying the probability of anomaly occurrence by a preset maximum safety margin coefficient yields the reduction in basic torque corresponding to the current risk level. When the probability of anomaly occurrence is zero or extremely low, the reduction in basic torque is close to zero, and the upper limit of the dynamic torque limit curve only needs to retain a small safety margin for normal drilling above the estimated formation equivalent resistance torque. As the probability of anomaly occurrence increases, the reduction in basic torque increases synchronously, causing the upper limit of dynamic torque to move closer to the estimated formation equivalent resistance torque, thus limiting overload cutting by the drill bit. The type of anomaly symptom modifies this calculation process, with different types of anomaly symptom corresponding to different additional adjustment coefficients. For example, when the anomaly symptom type is "pre-sticking," the additional adjustment coefficient is higher, further tightening the upper limit of torque; when the anomaly symptom type is "sticking signs," the focus is on reducing the upper limit of rotational speed, and the reduction in the upper limit of torque can be relatively wider. Multiply the base torque reduction by the additional adjustment coefficient corresponding to the abnormal symptom type to obtain the final dynamic torque reduction.

[0079] Using the absolute torque upper limit value of the equipment's capacity boundary as the starting point for adjustment, the final dynamic torque reduction amount is subtracted downwards to obtain the current torque upper limit value. This torque upper limit value also needs to be compared with the estimated formation equivalent resistance torque plus a minimum safe drilling margin, and the larger of the two is taken to avoid the torque upper limit being excessively reduced to below the level required for normal cutting, causing the drilling rate to drop to zero. The value determined after the above comparison is the torque upper limit value at the current moment.

[0080] A series of torque upper limit values ​​calculated consecutively in adjacent control cycles are connected sequentially along the time axis to form a dynamic torque limit curve that varies with time. Each data point on this curve corresponds to the torque upper limit decision result for one control cycle.

[0081] During the generation of the dynamic torque limit curve, a preset slope constraint is applied to the rate of change of the dynamic torque limit curve. The slope constraint has two independent limits: the maximum permissible upward slope and the maximum permissible downward slope. Within each control cycle, the currently calculated upper limit torque value is compared with the actual upper limit torque value effective in the previous control cycle. The difference between the two is calculated and divided by the control cycle duration to obtain the rate of change of the upper limit torque value. When the rate of change is positive (i.e., the upper limit torque is increasing), if it exceeds the maximum permissible upward slope, the upper limit torque value of the current cycle is corrected to the previous cycle value plus the maximum permissible upward slope multiplied by the control cycle duration to limit the acceleration rate during the recovery process. When the rate of change is negative (i.e., the upper limit torque is decreasing), if its absolute value exceeds the maximum permissible downward slope, the upper limit torque value of the current cycle is corrected to the previous cycle value minus the maximum permissible downward slope multiplied by the control cycle duration to ensure a smooth decrease in the upper limit torque during the risk tightening process. The two sets of slope limits are pre-configured according to the drilling engineering design requirements and equipment capacity. For example, the maximum allowable upward slope is 2% of the rated torque per second, and the maximum allowable downward slope is 5% of the rated torque per second.

[0082] Thus, after performing the above calculations and constraints in each control cycle, a dynamic torque limit curve that conforms to the equipment capacity boundary, reflects real-time formation resistance, responds to risk level, and is limited by the rate of change is obtained and output, which can be directly applied in subsequent linkage control and command issuance steps.

[0083] The equipment capacity boundary is a comprehensive engineering constraint description of the range of torque that the top drive system and its associated rotating systems can withstand and output under the premise of safe operation. The equipment capacity boundary is not a single fixed value, but a set of safe operating areas defined by multiple parameters and relationship curves. It clearly specifies the maximum allowable torque value output by the top drive system under different speeds, durations, and thermal conditions, and is a hard upper limit that cannot be exceeded when generating dynamic torque limit curves.

[0084] Rated torque, determined by the top drive system manufacturer during design and factory testing, represents the maximum torque that the top drive unit can output stably and continuously under specified cooling conditions and continuous operation. Rated torque is one of the most fundamental parameters within the equipment's capability limits. Under normal, stable drilling conditions, the upper limit of the dynamic torque limit curve typically will not exceed the rated torque for extended periods to ensure the design life and operational reliability of the top drive system.

[0085] Short-duration torque is the upper limit of torque that a top drive system is allowed to output briefly exceeding its rated torque for a limited period of time. The value of short-duration torque is usually higher than the rated torque by a certain percentage, such as 120% or 150% of the rated torque, but it corresponds to a strict maximum permissible duration, such as 30 seconds or 60 seconds. Short-duration torque provides a buffer for the top drive system when encountering short-term impact loads such as sudden hardening of the ground, allowing the torque to rise momentarily without immediately triggering limiting. However, after exceeding the permissible duration, the torque limiting strategy must intervene.

[0086] The thermal limit is the current maximum allowable torque limit calculated based on the real-time temperature or temperature rise model of the main heat-generating components of the top drive unit and rotating system, such as motor windings, gearbox lubricating oil, and bearings. When the top drive system operates under high load for an extended period, the temperature of the heat-generating components gradually rises, and the thermal limit will gradually decrease as the temperature increases to protect the equipment from insulation aging, lubrication failure, or mechanical damage caused by overheating. The thermal limit is a parameter in the equipment's capability boundary that dynamically changes with the operating conditions.

[0087] The speed-torque envelope relationship is a characteristic curve or envelope region showing how the output torque capability of a top drive unit changes with rotational speed. In the low-speed range, the top drive system can typically output higher torque, a capability limited by the maximum current or maximum hydraulic pressure. In the high-speed range, limited by the constant power characteristics of the drive motor or the maximum speed of the hydraulic motor, the maximum allowable output torque decreases as the speed increases. The speed-torque envelope relationship incorporates these speed-related torque limitations into the equipment's capability boundary in the form of a lookup table or curve function.

[0088] The formation equivalent resistance torque model is an engineering calculation model that estimates the torque required for the drill bit to cut rock, overcome friction, and carry drill cuttings at the bottom of the well, based on current drilling parameters and formation conditions. This model comprehensively considers factors such as drill bit type, formation lithology, wellbore structure, drilling fluid properties, and real-time drilling parameters, outputting an equivalent resistance torque estimate. During normal drilling, this estimate serves as an important reference for determining the lower limit of torque and optimizing the upper limit of torque; when the model's estimate deviates significantly from the measured torque, it indirectly reflects possible abnormal working conditions downhole.

[0089] The dynamic torque limit curve is a trajectory line that updates in real time with continuous time changes, representing the upper limit of the allowable torque. It is not a pre-set fixed line, but rather a curve recalculated and generated in each control cycle based on equipment capacity boundaries, the equivalent drag torque model of the formation, the probability of anomalies, and the type of anomaly symptoms. This curve accurately provides the maximum allowable torque value of the top drive system at the current moment and within a future time period, achieving adaptive and real-time safety constraints.

[0090] The control cycle refers to the fixed time interval in the entire dynamic torque limiting and linkage control process, from the initial data acquisition and risk assessment to the generation of the dynamic torque limit curve and the issuance of control commands, and then to the next recalculation. The duration of the control cycle is usually matched with the computing power and response requirements of the edge computing unit, and is set between tens of milliseconds and hundreds of milliseconds to ensure a rapid closed-loop response to changes in drilling conditions and risk events.

[0091] The preset slope constraint is a restriction on the maximum rate of change allowed for the dynamic torque limit curve per unit time. The slope constraint specifies the upper limits of the rising and falling slopes of the torque limit value between adjacent control cycles or per unit time, to avoid abrupt changes in the torque limit parameters. For example, when a high-risk situation is detected requiring a reduction in the torque limit, the reduction will be gradual and step-by-step according to the set maximum falling slope, rather than an instantaneous drop from full load to a certain low limit, thus preventing secondary impacts on the drill string and top drive transmission system caused by a sudden tightening of the torque limit.

[0092] S4. When the probability of this anomaly occurring exceeds a preset threshold, the current upper limit value of the dynamic torque limit curve is lowered, the upper limit of the rotational speed of the top drive system is reduced, and the drill string lifting speed or drilling pressure is limited. At the same time, a coordinated control signal is sent to the well control safety system to trigger the well control related devices to enter the pre-emptive state. The specific implementation process is as follows: S40. Determine if the probability of an anomaly exceeds a preset threshold. Within each control cycle, the trained TCN-Transformer model outputs an anomaly probability value after completing a risk assessment. This probability value is compared to a preset threshold. The preset threshold is pre-set based on the risk tolerance of the drilling area and historical data statistics, for example, 0.8. If the output anomaly probability is less than the preset threshold, the current drilling condition is considered to be within a controllable and normal range. No linkage control action is triggered within this control cycle, and the top drive system continues to operate in conventional control mode. If the anomaly probability is greater than or equal to the preset threshold, a high risk is identified, and the subsequent multi-parameter linkage intervention steps are immediately initiated. The comparison process is completed in the edge computing unit, taking only microseconds to ensure real-time triggering of the judgment.

[0093] S41. Lower the current upper limit value of the dynamic torque limit curve. When the probability of an anomaly exceeds a preset threshold, the dynamic torque limit curve, generated in the previous step, contains a torque upper limit value that changes over time for each data point on the curve. The torque upper limit value corresponding to the current control cycle is a lowered limit value calculated by integrating the equipment capacity boundary, the formation equivalent resistance torque model, the probability of anomaly occurrence, and the type of anomaly symptom. In this step, the upper limit parameter of the top drive torque that is actually effective in the current control loop is forcibly updated to the current upper limit value on the dynamic torque limit curve. This update operation is completed by writing the new torque upper limit value to the control parameter interface of the top drive device. After the torque upper limit value is lowered, the torque control logic of the top drive system will immediately perform closed-loop adjustment with the new upper limit value as the boundary. Once the real-time measured top drive torque approaches or touches the lowered upper limit value, the top drive device automatically reduces the output current or hydraulic pressure to reduce the torque, thereby avoiding the risk of drill bit overload and stuck drill due to torque exceeding the limit.

[0094] S42. Reduce the upper limit of the rotational speed of the top drive system. Synchronously with the reduction of the upper limit of torque, a reduced upper limit command for the rotational speed is generated in this step. The reduction in the upper limit of the rotational speed is determined by the probability of an anomaly and the type of anomaly symptom. The higher the probability of an anomaly, the larger the adjustment coefficient for the upper limit of the speed. If the anomaly symptom type is "pre-sticking" or "sticking signs," the reduction in the upper limit of the speed is relatively large to reduce the disturbance of the drill string rotation to the wellbore. If the anomaly symptom type is "mechanical imbalance," the reduction in the upper limit of the speed is relatively conservative, possibly limiting the upper limit of the speed to within 1.1 times the current measured speed to prevent excessive speed from aggravating mechanical vibration, while not completely stopping the rotation. The reduced upper limit of the rotational speed is used as the maximum allowable speed of the current speed control loop, and the top drive unit adjusts the actual rotational speed below this upper limit.

[0095] S43. Limit drill string lift speed or drilling pressure. This step is performed simultaneously with torque and speed limits to constrain the axial motion parameters of the drill string. The limiting method selects one or both of drill string lift speed and drilling pressure based on the type of anomaly and well conditions. If the current anomaly indicates a risk of wellbore instability due to suction or agitation pressure, drill string lift speed is limited first. A maximum allowable lift speed command is sent to the drilling winch control loop, setting the upper limit of the lift speed to a low preset safety value, such as 0.1 meters per second. If the anomaly indicates abnormal drill bit cutting or excessive drilling pressure fluctuations, drilling pressure is limited first. A reduced upper limit of drilling pressure is sent to the top drive's drilling pressure control loop or the automatic drill bit feed system. The reduction in the upper limit of drilling pressure can be determined based on the real-time fluctuation intensity of drilling pressure, and the reduction ratio is usually between 10% and 30% of the current drilling pressure. After the command to limit the drill string lifting speed or drilling pressure is issued, the winch or top drive will comply with the new upper limit in automatic mode. If the drilling site is in manual operation mode, the limit will also serve as an audible and visual alarm to prompt the operator to adjust the handle range.

[0096] S44. Send a coordinated control signal to the well control safety system to trigger the well control-related devices to enter a pre-emptive state. Simultaneously with the execution of the aforementioned torque, speed, and load reduction control commands, a coordinated control signal is generated. The method for generating the coordinated control signal is to construct a standardized message data packet. This message data packet includes a message identifier, timestamp, current anomaly probability value, anomaly symptom type, a summary of the linked control actions, and suggested preventative actions for the well control-related devices. Specific suggested preventative actions include a pre-closing indication for the annular blowout preventer, a self-check of the hydraulic system's energy storage for the gate blowout preventer, a suggested preset opening degree for the choke manifold choke valve, and a self-start indication for the standby pump in the mud circulation system. This coordinated control signal is sent to the well control safety system's communication interface via a soft bus using a publish-subscribe or point-to-point method. After receiving and analyzing the collaborative control signal, the well control safety system automatically or manually pressurizes the blowout preventer's accumulator to its maximum operating pressure according to pre-configured risk response rules. It also adjusts the throttle valve opening to the preset waiting throttle position, starts the auxiliary mud pump, and switches the valve assembly to standby mode. The mud level monitoring system increases its sampling frequency to detect minor overflows or leaks. Once the well control devices enter a preventative state, the start-up delay of well control operations is significantly reduced, enabling rapid well shut-in or the establishment of a circulating kill system should the abnormal situation worsen.

[0097] After all the above steps have been completed, the entire coordinated intervention is finished. The top drive system and the associated downhole hoisting and well control safety systems have been uniformly adjusted to a safe constraint state within milliseconds. Subsequently, the control process enters the continuous monitoring phase. In subsequent control cycles, if the probability of an anomaly occurs continues to be higher than the preset threshold, the above-mentioned restrictions will remain in place; if the probability of an anomaly occurs falls below the preset release threshold, the rollback strategy process will be initiated to gradually remove the various restrictions.

[0098] The preset threshold is a pre-defined decision boundary value stored in the control strategy configuration, used to distinguish between normal operating conditions and abnormal operating conditions that require triggering linkage intervention. This threshold is a constant between 0 and 1, with a typical configuration value in the range of 0.6 to 0.9. When the probability of anomaly occurrence output by the trained TCN-Transformer model is greater than or equal to this preset threshold, the linkage control process is activated; when the probability of anomaly occurrence is lower than this threshold, the top drive system maintains normal control status and does not impose additional restrictions on parameters such as torque and speed.

[0099] The upper limit of rotational speed is the highest allowed rotational speed of the top drive system during operation. It is a control command parameter, distinct from the actual measured rotational speed in the target multi-source sensor data. The upper limit of rotational speed directly constrains the maximum output speed of the top drive unit; under no circumstances will the speed control loop of the top drive system allow the actual rotational speed to exceed this limit. During normal drilling, the upper limit of rotational speed is set according to the drilling design parameters; when the linkage control is activated, this limit is actively reduced to minimize downhole disturbance caused by drill string rotation.

[0100] Drill string lift-up speed refers to the linear speed of the drill string along the wellbore axis when the drilling winch lifts or lowers the drill string. The magnitude of the drill string lift-up speed directly affects bottom hole pressure fluctuations and the friction between the drill string and the wellbore. Limiting the drill string lift-up speed under abnormal operating conditions is to prevent bottom hole suction pressure or agitation pressure caused by rapid drill string lifting and lowering, thus avoiding these pressure disturbances from becoming additional factors that could collapse the wellbore or induce a well kick.

[0101] Drill pressure (DP) refers to the axial pressure exerted by the drill bit on the rock at the bottom of the well. It is typically provided by the weight of the drill string itself and controlled and regulated by the top drive or winch. The magnitude of DP determines the depth to which the drill bit's cutting teeth penetrate the rock and the rock-breaking efficiency. Limiting DP in a linkage control system involves reducing the axial load applied to the drill bit through control commands, thereby reducing the torque demand at the drill bit and mechanical disturbance at the bottom of the well. This, combined with measures such as lowering the torque limit, helps to suppress the escalation of abnormalities.

[0102] The well control safety system is an independently operating safety control system at the drilling site, specifically responsible for monitoring and handling major safety incidents related to wellbore pressure control, such as well kicks and blowouts. The well control safety system connects to well control-related devices such as blowout preventer assemblies, choke manifolds, kill manifolds, and mud circulation systems, and has the capability to automatically or manually trigger well control operations. In this technical solution, the linkage control sends coordinated control signals to the well control safety system to preemptively put the well control safety system into a state of alert when high-risk signs are detected on the top drive side.

[0103] The collaborative control signal is a standardized instruction message generated during the automatic control process and directed to other independent control systems. This signal does not directly force the hardware at the execution end to operate; instead, it transmits information such as risk level, type of abnormal symptom, and suggested collaborative actions, informing the well control safety system that an anomaly has been detected on the top drive side and recommending that the well control safety system proactively raise the prevention level. The collaborative control signal is published via a soft bus or industrial communication protocol and includes data fields such as timestamp, risk level, and blowout preventer pre-action prompts.

[0104] Well control-related devices (BCDs) are a collective term for various hardware devices used at the drilling site to prevent, detect, and control abnormal wellbore pressure. Typical BCDs include annular blowout preventers (BOPs), gate blowout preventers (BOPs), choke manifolds, kill manifolds, mud pumps, liquid-gas separators, and mud tanks. Upon receiving a coordinated control signal, these BCDs enter a pre-emptive state, such as ensuring the annular BOP core is pre-charged to the correct pressure, the choke valve is preset to the correct opening, and the standby mud pump is automatically activated and ready to go, in order to shorten the actual response time when a well control event actually occurs.

[0105] The "precautionary mode" is a highly prepared operating mode that the well control safety system and related devices enter in advance before an actual well control event occurs. In this mode, the blowout preventer hydraulic accumulator pressure is maintained at its maximum operating value, key valves are switched to specific standby positions, weighting and plugging materials for the mud circulation system are ready, and operators receive early warning notifications. The precautionary mode does not affect the continuous execution of normal drilling operations, but it ensures that well control operations such as shut-in or choke circulation can be completed in the shortest possible time if needed.

[0106] The following combination Figure 2 and Figure 3 The components of the method of the present invention will be described in detail.

[0107] in, Figure 2 It illustrates the relationship between the field sensor cluster, edge computing unit, and actuators, demonstrating the complete process from data acquisition and intelligent analysis to control execution. Specifically, Figure 2This paper showcases the overall architecture of a dynamic torque limiting and early warning linkage control method for top drive systems, clearly depicting the complete data flow path from multi-source sensor data acquisition at the drilling site, edge-side intelligent analysis, to control command issuance and execution feedback. On the far left of the architecture, various sensors deployed at the drilling site constitute the data sensing layer. These sensors collect physical quantities reflecting real-time drilling conditions, specifically including top drive torque and rotational speed, hook load and drilling pressure, mud pump pressure, and drill string vibration and acceleration signals. Figure 2 The data is categorized by torque / speed, hook load / drilling pressure, mud pump pressure, and vibration / acceleration, respectively. These physical quantities constitute the data source for all subsequent intelligent analyses. The raw data collected by these sensors is not directly transmitted to the analysis unit, but is instead uniformly integrated into a unified data and command channel comprised of a unified semantic point table management system and a distributed soft bus provided by the industrial operating system. Figure 2 The “Unified Semantic Point Table & Soft Bus (Data / Command Unified Channel)” indicated in the middle is a channel that converts multi-source heterogeneous sensor data of different protocols and formats into engineering data identified by standard semantic objects according to a pre-configured semantic-point mapping file, forming target multi-source sensor data, and transmitting it to the edge computing unit at high speed in the form of a standardized data stream.

[0108] Figure 2 The edge computing unit (OS) in the middle is the intelligent decision-making hub of the entire methodology, internally running two closely cooperating processing threads. The first thread is risk identification, corresponding to... Figure 2 The first main line is the processing box for "Risk Identification: TCN-Transformer Sliding Window Input-Output Abnormal Probability p, Symptom c". In this main line, the time series of target multi-source sensor data is divided into multiple time windows according to a preset sliding window length. After feature extraction, a feature tensor is formed and input into the trained TCN-Transformer model. The TCN layer captures local temporal features, and the Transformer layer extracts global dependency features across channels and across windows. Finally, it outputs the probability of anomaly occurrence p and the type of abnormal symptom c. Here, the symbol p represents a continuous value between 0 and 1, used to quantify the probability of an abnormal condition existing within the current time window, and the symbol c represents the specific risk category label identified by the model, such as "adhesion" or "early signs of stuck drill". The second main line is dynamic torque limit calculation and linkage control generation, corresponding to Figure 2The system comprises two processing boxes: "Dynamic Torque Limit Calculation Equipment Capacity Boundary + Formation Equivalent Resistance Torque + Slope Constraint" and "Linked Control and Retreat Torque / Speed / Load Limit + Well Control Collaborative Confidence / Progressive Retreat t". In the dynamic torque limit calculation, the equipment capacity boundary is first obtained from the rated torque, short-time torque, thermal limit value, and speed-torque envelope relationship of the top drive system and rotational system. Simultaneously, the formation equivalent resistance torque model is corrected online based on the target multi-source sensor data, including speed, drilling pressure, mud pump pressure, well depth, friction, and mud parameters. Then, combining the anomaly probability p and anomaly symptom type c, the upper limit value of torque changing with time is calculated in each control cycle, generating a dynamic torque limit curve L(t). A preset slope constraint is applied to the rate of change of this curve. The symbol L(t) represents the upper limit value function of torque with time t as the independent variable. In the linkage control and rollback, when the probability of an anomaly p exceeds a preset threshold, a coordinated control command including torque limiting, speed limiting, and load limiting is automatically generated, and a control command confidence level based on p is attached to the command so that the operator can understand the reliability of the intervention. When the abnormal situation is resolved, the parameters are restored in a gradual manner according to the rollback strategy. The symbol t here represents the predetermined time constant used in the rollback process.

[0109] Figure 2 The right side of the diagram illustrates the path from command issuance to execution feedback. The "commands / settings / coordination" generated by the edge computing unit are sent directionally to the top drive controller / well control (BOP / cycle) of the field actuators via a unified semantic point table or soft bus (unified data / command channel). Specifically, the top drive unit and its controller receive torque limits and rotational speed adjustment commands; the drilling hoisting system receives commands limiting drill string hoisting speed or drilling pressure; and the well control safety system receives coordinated control signals to trigger well control-related devices such as blowout preventers and mud circulation systems to enter a pre-emptive state. Simultaneously, the actuators transmit their own events / status, along with actual torque, speed, and load limits, as feedback / status data via the soft bus back to the edge computing unit (OS) for dynamic calculation and retrospective auditing in the next control cycle. Furthermore, Figure 2 The "HMI / Remote Monitoring (Status Panel / Alarm / Parameter) Data" at the bottom is connected to the entire system through the data access and collaboration (top drive controller / well control system) interface. It presents standardized engineering data, dynamic torque limit curve L(t), probability of anomaly p, type of anomaly c, control command log and execution receipt, and other multi-dimensional information to the on-site team and remote management personnel in the form of status panel, alarm list and parameter curve, thereby completing a complete data closed loop from data perception, risk identification, linkage intervention to human-machine monitoring.

[0110] in, Figure 3The flowchart presents the complete logical process of risk assessment and linkage control in the dynamic torque limiting and early warning linkage control method for top drive systems. Starting from data input, it forms a continuous processing path through intelligent analysis, decision judgment, linkage execution, and anomaly recovery. The process begins with data acquisition and preprocessing. Multiple sensors deployed at the drilling site collect multi-source sensor data in real time, such as top drive torque, rotational speed, hook load, drilling pressure, mud pump pressure, and drill string vibration signals, to reflect the real-time drilling conditions. These raw data frames are aggregated to the edge computing unit through a distributed soft bus. After protocol type identification and register address space parsing, they are converted into engineering data identified by standard semantic objects according to the pre-configured semantic-point mapping file, resulting in target multi-source sensor data. Subsequently, the time series of the target multi-source sensor data is divided into multiple time windows according to a preset sliding window length. At least one feature is extracted from the data in each time window, including time domain statistics, envelope, frequency domain energy, spectral peak stability, and amplitude variation rate, to form a feature tensor. After preprocessing, the feature tensor is stored in the sliding time window. Next, the process enters the analysis phase of the risk identification mini-model TCN-Transformer, i.e., the trained TCN-Transformer model. Feature tensors are input into this model; the TCN layer captures local temporal features, and the Transformer layer extracts global dependency features across channels and windows. Based on the local temporal features and global dependency features, the model outputs the probability of anomaly occurrence p and the type of anomaly symptom c. Here, p represents a continuous value between 0 and 1, used to quantify the probability of abnormal conditions in the drilling process within the current time window. c represents the specific risk category label identified by the model, such as "sticking" or "precursor to stuck drill bit." The process then determines whether p is greater than or equal to θ, i.e., whether the probability of anomaly occurrence p is greater than or equal to a preset threshold θ. θ represents a pre-set risk judgment threshold value for triggering linkage control, a constant between 0 and 1, with a typical configuration value in the range of 0.6 to 0.9. When the judgment result is p less than θ, meaning the probability of an anomaly does not exceed the preset threshold, the process moves to the normal control branch. The top drive system maintains normal torque, speed, and drilling pressure control states without triggering additional limiting measures. Simultaneously, the operating parameters under this normal control state are recorded in the execution feedback and log for archiving and subsequent auditing. When the judgment result is p greater than or equal to θ, meaning the probability of an anomaly exceeds the preset threshold, the process enters two parallel control branches.The first branch is the dynamic torque limit calculation L(t). In this branch, the equipment capacity boundary is first obtained by the rated torque, short-time torque, thermal limit value and speed-torque envelope relationship of the top drive system and the rotating system. At the same time, the formation equivalent resistance torque model is corrected online based on the speed, drilling pressure, mud pump pressure, well depth, friction and mud parameters in the target multi-source sensor data. Then, combined with the probability of anomaly occurrence p and the type of anomaly c, the upper limit value of torque changing with time is calculated in each control cycle, generating the dynamic torque limit curve L(t), and a preset slope constraint is applied to the rate of change of the curve. The symbol L(t) represents the torque upper limit value function with time t as the independent variable. The second branch is the execution of coordinated control. Based on the generated dynamic torque limit curve L(t), the current upper limit of torque is lowered, reducing the upper limit of the top drive system's rotational speed and imposing restrictions on drill string lifting speed or drilling pressure. Simultaneously, a coordinated control signal is sent to the well control safety system to trigger well control-related devices such as the blowout preventer and mud circulation system to enter a pre-emptive state. These control commands all include a confidence level based on the probability of anomaly occurrence p. All commands, responses, and status data during the coordinated control execution process are also recorded in the execution feedback and log. After the coordinated control execution, the process enters the anomaly resolution judgment node, which continuously monitors whether the probability of anomaly occurrence p has fallen below the preset release threshold and remained stable to determine whether the abnormal condition has been resolved. If the anomaly persists, the process returns to continue executing the dynamic torque limit calculation L(t) and linkage control, maintaining all limiting measures. If the anomaly has subsided, the process switches to the rollback strategy branch. According to the pre-set rollback strategy, parameters such as torque limits and rotational speed are gradually increased to normal levels. The rollback time constant is represented by the symbol t, and its value is adjusted between 5 and 120 seconds based on the well section and operating condition template. If an increase in the probability of anomaly occurrence p is detected again during the rollback process, the rollback is immediately paused or the control strategy is tightened again. The status data of the rollback process is also continuously written to the execution feedback and log. Finally, when all parameters have smoothly returned to normal, the process rejoins the normal control branch to continue continuous monitoring of the drilling process. The entire flowchart graphically illustrates the logical closed loop from data monitoring, anomaly identification, linkage intervention to smooth recovery.

[0111] The method in this embodiment includes: 1) Sensor clustering and data semantic management are implemented to acquire target multi-source sensor data. Multiple sensors deployed at the drilling site are used to collect key parameters of the top drive system and drilling process in real time, including but not limited to top drive torque, rotational speed, hook load, drilling pressure, mud pump pressure, and drill string vibration signals. These raw data frames collected by the sensors are aggregated to the edge computing unit via a distributed soft bus provided by the industrial operating system. In the edge computing unit, protocol type identification and register address space parsing are performed on these raw data frames to identify the industrial protocols they follow, such as TCP / IP, MQTT, Profinet, Profibus, CAN, OPC UA, Modbus, and HART. Subsequently, according to a pre-configured semantic-point mapping file, the observations corresponding to the key parameters in the raw data frames are converted into engineering data identified by standard semantic objects, thereby achieving standardized fusion and high-speed transmission of multi-source data to obtain uniformly managed target multi-source sensor data.

[0112] 2) A trained TCN-Transformer model, i.e., an edge-side risk identification model, is built and run in the edge computing unit to perform real-time risk assessment on the time series of target multi-source sensor data. This model consists of a TCN layer and a Transformer layer. The specific processing is as follows: the target multi-source sensor data is divided into multiple time windows according to a preset sliding window length (e.g., 128 data points). At least one feature is extracted from the data in each time window, including time-domain statistics, envelope, frequency-domain energy, spectral peak stability, and amplitude variation rate, to obtain a feature tensor. Subsequently, the TCN layer in the trained TCN-Transformer model captures local temporal features of the feature tensor, and the Transformer layer extracts global dependency features across channels and across windows. This enables high-precision early identification of weak signs such as "sticking, slippage, drill jamming precursors, and mechanical imbalance" under the limited computing resources of the edge device, and outputs the current probability of anomaly occurrence and the type of anomaly sign.

[0113] 3) Based on the risk assessment results, perform dynamic torque limit calculations to generate a dynamic torque limit curve L(t). First, obtain the equipment capacity boundary of the top drive system. This boundary is determined by the rated torque, short-term torque, thermal limit value, and speed-torque envelope relationship of the top drive system and the rotary system, representing the limits that the top drive drive device and drill string can withstand. Simultaneously, based on the target multi-source sensor data (speed, drilling pressure, mud pump pressure, well depth, friction, and mud parameters), the formation equivalent resistance torque model is corrected online to estimate the torque required for cutting under current formation conditions. Within each control cycle, based on the equipment capacity boundary and the corrected formation equivalent resistance torque model, combined with the probability of anomaly occurrence and the type of anomaly symptoms, calculate the upper limit value of torque changing over time, thereby generating the dynamic torque limit curve L(t). This curve applies a preset slope constraint to the rate of change to avoid excessive tightening causing secondary impacts, and is updated in real time to ensure that the torque limit is neither too conservative, affecting drilling efficiency, nor too lagging, causing missed opportunities for safety intervention.

[0114] 4) When the probability of an anomaly output by the risk assessment model exceeds a preset threshold and triggers an anomaly event, a series of coordinated control actions are immediately executed. Specifically, the current upper limit of the dynamic torque limit curve is lowered, the upper limit of the rotational speed of the top drive system (i.e., speed limiting control) is reduced, and the drill string lifting speed or applied drilling pressure (i.e., load limiting control) is limited to reduce downhole impact and prevent risk escalation. The linkage control logic sends coordinated control signals to the well control safety system through the soft bus interface of the industrial operating system to trigger well control-related devices (such as blowout preventers, mud circulation systems, etc.) to enter a pre-emptive state, ensuring drilling safety from multiple levels. The entire process requires no manual intervention and can be completed within milliseconds, achieving immediate response to abnormal signs and multi-system linkage intervention. To ensure the reliability of the control link, a unified message format and priority queue mechanism are adopted. Control messages have higher delivery priority and expiration logic, and packet loss retry and duplication suppression are achieved in conjunction with ACK / NAK and idempotent flags. Once the execution end fails to return ACK within the specified time limit, it enters a preset safety degradation state, continues to work using a conservative factory-set safety curve, and simultaneously issues alarms.

[0115] 5) Implement control command confidence levels, i.e., confidence parameters and rollback strategies. All control commands issued using this method are accompanied by a confidence level based on an anomaly probability assessment, allowing the drilling control system and operators to understand the reliability and severity level of the automatic intervention. Once the anomaly is detected and resolved (i.e., the anomaly probability drops to a safe level), the previously implemented torque, speed, and load limiting measures are gradually lifted according to a pre-set rollback strategy. This involves progressively increasing torque limits and rotational speeds to normal levels. The rollback time constant can be adjusted between 5 and 120 seconds based on the well section and operating conditions to avoid sudden changes in control parameters that could trigger new shocks. If an increased risk trend is detected again during the recovery process, the rollback is promptly paused or the control strategy is tightened again, ensuring a closed loop from "real-time perception" to "active intervention" to "smooth recovery." By adding control command confidence markers to commands and setting rollback mechanisms, transparency and safety redundancy in control decisions are guaranteed.

[0116] In another embodiment, a top-drive torque limiting and early fault linkage control method running in an industrial IoT operating system environment is presented. This method, without limiting specific hardware configurations, uses the OS's distributed soft bus and unified semantic point table management as its foundation to achieve parsing, semantic mapping, and real-time control of raw data in common industrial protocols for top drives. The overall idea is to unify multi-source heterogeneous signals from the well site into standard engineering objects at the semantic layer. After real-time identification of risk signs by a lightweight edge-side model, a dynamic torque limit curve is generated according to the comprehensive boundary between the equipment and the formation, and then linked with speed limiting, load limiting, and well control collaborative strategies for execution. Specifically, the method includes the following steps: 1) Complete the disk writing and distribution of raw protocols to standard engineering semantics. Interact with raw data frames from the top drive and drilling process, identify their protocol type and register / node address space, assign unified engineering attributes to each observation and control quantity, and generate a semantic-point mapping file organized hierarchically as "standard semantic object—combined object—protocol scan block—configuration distribution". Standard semantic objects define attributes such as top drive torque, rotary speed, hook load, drill pressure, riser pressure, drill string vibration signal, and torque limit setting. Combined objects aggregate these attributes according to "state set / control set", and protocol scan blocks group consecutive address ranges and poll them in a block-based manner. After the mapping file configuration is distributed, raw protocol data is written to the raw buffer during runtime, and a standard semantic buffer is generated synchronously according to the mapping rules. All subsequent steps are only for standard semantic access.

[0117] 2) The sampling rate of each channel of the target multi-source sensor data is aligned and self-synchronized, abnormal timestamps and duplicate frames are removed, and a quality bit and time identifier are attached to each sample to form a sliding window sequence of length N, typically with 128 to 256 points, corresponding to a sampling domain of 256 Hz to 1000 Hz. At the same time, lightweight features such as time-domain statistics, envelope, frequency-domain energy, spectral peak stability, and amplitude variation rate are generated and written into the feature cache in the form of feature tensors for edge inference consumption.

[0118] 3) Load a lightweight, trained TCN-Transformer model into the OS's edge computing environment, and perform millisecond-level inference using the aforementioned feature tensors. The TCN layer is responsible for capturing local temporal textures and short-term mutations, while the Transformer layer extracts global dependencies across channels and windows. The model outputs the anomaly occurrence probability p, the anomaly symptom type c, and the intensity index. The inference results are encapsulated as risk event messages and published to the soft bus, carrying the source, timestamp, quality bit, and signature verification.

[0119] 4) Within each control cycle, the dynamic torque limit curve L(t) is calculated in real time by integrating the equipment capacity boundary and the formation equivalent resistance torque model. The equipment capacity boundary is derived from the rated torque, short-term torque, thermal limit value, and speed-torque envelope of the top drive and rotational system. The formation equivalent resistance torque model is corrected online based on the current speed, drilling pressure, standpipe pressure, well depth, friction, and mud parameters. The allowable torque upper limit is given in the form of a continuous-time curve, and a slope constraint is applied to its rate of change. When p exceeds the threshold or c exhibits specific symptoms, the curve is flexibly lowered in the current segment according to the strength coefficient. The flexible lowering coefficient can be configured in stages between 5% and 30%, and the source and duration of the lowering are recorded. L(t) is issued as a standard semantic object and can be directly interpreted by the execution side as a torque limit reference and a time-varying boundary.

[0120] 5) Receive L(t) and risk level, package control commands such as torque limiting, speed limiting, and load limiting, and send them to the execution end via the OS soft bus. Simultaneously, issue coordinated triggering quantities to relevant well control subsystems in a publish / subscribe format, such as maintaining stable speed, increasing cuttings removal volume, and providing pump volume coordination suggestions. The execution end transmits back the actual L(t) segment and key process quantities in each cycle for closed-loop verification and retrospective auditing.

[0121] 6) When the probability of an anomaly occurs remains below the release threshold, the rollback strategy smoothly restores torque, speed, and load limits to the baseline according to a set time constant (5 to 120 seconds, adjusted according to the well section and operating condition template). If an upward trend in risk is detected again during the recovery process, the rollback is immediately paused or the control strategy is tightened again. If the link is unstable or execution is abnormal, the safety degradation takes effect immediately and a remote alarm is issued.

[0122] During an actual drilling operation, the system detected a combination of weak signs: a slow increase in torque, a synchronous increase in hook load, and a decrease in rotational speed. The trained TCN-Transformer model gave an anomaly probability p≈0.8 and an anomaly type c="adhesion" within a single inference cycle. The dynamic torque limit curve was then adjusted downwards to a moderate level. The linkage control moderately tightened the upper limit of rotational speed and the lifting rate, and issued cuttings removal and stabilization suggestions to the well control side. After confirmation by the executor, the real-time torque stabilized below the adjusted L(t). After tens of seconds, the anomaly probability dropped below the release threshold, and the system smoothly recovered to the original curve according to the predetermined time constant. No shocks or secondary fluctuations occurred throughout the process. Risk events, curve changes, control commands, and execution receipts during the period were all recorded and documented using standard semantic logging.

[0123] Through the above implementation process, this method uses semantic mapping and a soft bus as a unified foundation, a lightweight edge-side model as a risk identification tool, a dynamic torque limit curve as a constraint, and a message bus and reliable execution as the control path, forming a continuous main line from data perception to coordinated handling and then to strategy rollback. Without being heavily reliant on specific hardware configurations, it can adapt to the operational needs of multiple protocols, scenarios, and roles, significantly improving the proactive perception, coordinated handling, and auditable and maintainable capabilities of top drive operations.

[0124] The purpose of this invention is to provide a dynamic torque limiting and early warning linkage edge control method for top drive systems. This method constructs a comprehensive control system encompassing "data perception—risk identification—torque limiting linkage—strategy rollback," thereby addressing the problems of static lag, reliance on manual intervention, and lack of linkage between early warning and control in existing top drive torque control strategies. This invention utilizes a platform running on an industrial operating system, applying distributed soft bus and unified semantic point table management to drilling site data acquisition and control command transmission. This constructs a high-speed and stable data channel, enabling unified management and real-time sharing of target multi-source sensor data. With this architecture, this invention can intelligently analyze target multi-source sensor data at the edge, promptly identify risks and output anomaly probabilities and anomaly symptom types through a trained TCN-Transformer model, dynamically adjust dynamic torque limit curves and related control parameters, and smoothly roll back based on control command confidence when anomalies are resolved. Ultimately, this improves the automation and safety level of the drilling process.

[0125] The dynamic torque limiting and early warning linkage control method provided by this invention significantly improves the safety, adaptability, and response speed of the top drive system under complex drilling conditions. By using a trained TCN-Transformer model to identify risk signs in advance, the method can proactively take measures before the traditional fixed threshold passively triggers, achieving an early response to anomalies and significantly reducing the probability of accidents such as stuck pipe and well kicks. Compared with traditional solutions that rely solely on fixed torque limits and manual monitoring, this invention can adjust the torque upper limit in real time according to the drilling status, avoiding the problems of excessively high torque settings leading to equipment damage or excessively low torque settings affecting drilling efficiency, successfully achieving adaptive optimization of the torque control strategy.

[0126] Thanks to the distributed soft bus and unified semantic point table management mechanism provided by the industrial operating system, the solution of this invention has excellent engineering deployment and compatibility. Each step works collaboratively through unified data semantics and high-speed communication interfaces, allowing integration of the method without significant modifications to the existing drilling rig control system, thus reducing deployment costs and implementation difficulty. Simultaneously, the "data perception—risk identification—linkage control—strategy rollback" control loop constructed by this invention enables automated risk management capabilities at the drilling site. The entire process is executed cyclically within millisecond cycles, ensuring timely intervention from detection when the probability of anomalies exceeds limits, and supporting a smooth recovery of normal operations according to the rollback strategy after the risk is eliminated. This control method significantly improves the safety margin and stability of the top drive system, reduces reliance on manual decision-making, and enhances the overall efficiency of drilling operations.

[0127] In the above embodiments, although the steps are numbered S1, S2, etc., they are only specific embodiments given by the present invention. Those skilled in the art can adjust the execution order of S1, S2, etc. according to the actual situation. The scheme after adjusting the order is also within the protection scope of the present invention. It can be understood that in some embodiments, some or all of the above embodiments may be included.

[0128] like Figure 4 As shown, an embodiment of the present invention provides a dynamic torque limiting and early warning linkage edge control system 200 for a top drive system, which includes a data acquisition module 201, a risk assessment module 202, a curve generation module 203, and a linkage control module 204. The data acquisition module 201 is used to: acquire multi-source sensor data that reflects the real-time working conditions of drilling in real time, and perform unified semantic point table management to obtain target multi-source sensor data; The risk assessment module 202 is used to: use the trained TCN-Transformer model to perform risk assessment on the time series of target multi-source sensor data to obtain the probability of anomaly occurrence and the type of anomaly symptoms; The curve generation module 203 is used to generate a dynamic torque limit curve based on the probability of anomaly occurrence and the type of anomaly symptoms, combined with the rated performance parameters of the top drive system and the target multi-source sensor data. The linkage control module 204 is used to: when the probability of the anomaly exceeds the preset threshold, lower the current upper limit of the torque of the dynamic torque limit curve, reduce the upper limit of the rotation speed of the top drive system, and limit the drill string lifting speed or drilling pressure, while sending a coordinated control signal to the well control safety system to trigger the well control related devices to enter the prevention state in advance.

[0129] Optionally, in the above technical solution, the target multi-source sensing data includes at least one of top drive torque, rotational speed, hook load, drilling pressure, mud pump pressure, and drill string vibration signal; the multi-source sensing data includes raw data frames collected by multiple sensors; the data acquisition module 201 is specifically used to: aggregate the raw data frames to the edge computing unit through a distributed soft bus; perform protocol type identification and register address space parsing on the raw data frames to identify the protocol type corresponding to the raw data frames; and convert the observations corresponding to the target multi-source sensing data in the raw data frames into engineering data identified by standard semantic objects according to a pre-configured semantic-point mapping file, and use the engineering data as the target multi-source sensing data.

[0130] Optionally, in the above technical solution, the risk assessment module 202 is specifically used to: divide the time series of target multi-source sensing data into multiple time windows according to a preset sliding window length; extract at least one feature from the target multi-source sensing data in each time window, including time domain statistics, envelope, frequency domain energy, spectral peak stability, and amplitude variation rate, to obtain a feature tensor; input the feature tensor into a trained TCN-Transformer model, where the TCN layer in the trained TCN-Transformer model captures local temporal features, and the Transformer layer in the trained TCN-Transformer model extracts global dependency features across channels and across windows; and output the probability of anomaly occurrence and the type of anomaly symptoms based on the local temporal features and the global dependency features.

[0131] Optionally, in the above technical solution, the curve generation module 203 is specifically used to: obtain the equipment capability boundary of the top drive system, which is determined by the rated torque, short-time torque, thermal limit value, and speed-torque envelope relationship of the top drive system and the rotating system; based on the speed, drilling pressure, mud pump pressure, well depth, friction, and mud parameters in the target multi-source sensor data, correct the formation equivalent resistance torque model online, which is used to estimate the torque required for the drill bit to cut under the current formation conditions; within each control cycle, based on the equipment capability boundary and the corrected formation equivalent resistance torque model, combined with the probability of anomaly occurrence and the type of anomaly symptom, calculate the upper limit value of torque changing with time, thereby generating a dynamic torque limit curve, and apply a preset slope constraint to the rate of change of the dynamic torque limit curve.

[0132] It should be noted that the beneficial effects of the dynamic torque limiting and early warning linkage edge control system 200 for top drive systems provided in the above embodiments are the same as those of the dynamic torque limiting and early warning linkage edge control method for top drive systems described above, and will not be repeated here. Furthermore, the system provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the system can be divided into different functional modules according to the actual situation to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiments, and will not be repeated here.

[0133] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned dynamic torque limiting and early warning linkage edge control methods for top drive systems.

[0134] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned dynamic torque limiting and early warning linkage edge control methods for top drive systems.

[0135] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0136] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A dynamic torque limiting and early warning linkage edge control method for top drive systems, characterized in that, include: Real-time acquisition of multi-source sensor data reflecting real-time drilling conditions, and unified semantic point table management to obtain target multi-source sensor data; The trained TCN-Transformer model is used to conduct a risk assessment on the time series of multi-source sensor data of the target to obtain the probability of anomaly occurrence and the type of anomaly symptoms. Based on the probability of the anomaly and the type of the anomaly, combined with the rated performance parameters of the top drive system and the target multi-source sensor data, a dynamic torque limit curve is generated. When the probability of this anomaly exceeds a preset threshold, the current upper limit of the torque of the dynamic torque limit curve is lowered, the upper limit of the rotational speed of the top drive system is reduced, and the drill string lifting speed or drilling pressure is limited. At the same time, a coordinated control signal is sent to the well control safety system to trigger the well control related devices to enter the prevention state in advance.

2. The dynamic torque limiting and early warning linkage edge control method for top drive systems according to claim 1, characterized in that, The target multi-source sensing data includes at least one of top drive torque, rotational speed, hook load, drilling pressure, mud pump pressure, and drill string vibration signal; the multi-source sensing data includes raw data frames collected by multiple sensors; Real-time acquisition of multi-source sensor data reflecting real-time drilling conditions, and unified semantic point table management to obtain target multi-source sensor data, including: aggregating the original data frames to the edge computing unit through a distributed soft bus; The original data frame is subjected to protocol type identification and register address space parsing to identify the protocol type corresponding to the original data frame; according to the pre-configured semantic-point mapping file, the observations in the original data frame corresponding to the target multi-source sensing data are converted into engineering data identified by standard semantic objects, and the engineering data is used as the target multi-source sensing data.

3. The dynamic torque limiting and early warning linkage edge control method for top drive systems according to claim 1, characterized in that, Risk assessment of the time series composed of the target multi-source sensor data using a trained TCN-Transformer model includes: dividing the time series composed of the target multi-source sensor data into multiple time windows according to a preset sliding window length; extracting at least one feature from the target multi-source sensor data within each time window, including time-domain statistics, envelope, frequency-domain energy, spectral peak stability, and amplitude variation rate, to obtain a feature tensor; inputting the feature tensor into the trained TCN-Transformer model; capturing local temporal features by the TCN layer in the trained TCN-Transformer model; extracting global dependency features across channels and across windows by the Transformer layer in the trained TCN-Transformer model; and outputting the probability of anomaly occurrence and the type of anomaly symptom based on the local temporal features and the global dependency features.

4. The dynamic torque limiting and early warning linkage edge control method for top drive systems according to claim 1, characterized in that, Based on the probability of anomaly occurrence and the type of anomaly symptom, combined with the rated performance parameters of the top drive system and the target multi-source sensor data, a dynamic torque limit curve is generated. This includes: obtaining the equipment capability boundary of the top drive system, which is determined by the rated torque, short-term torque, thermal limit value, and speed-torque envelope relationship of the top drive system and the rotating system; based on the speed, drilling pressure, mud pump pressure, well depth, friction, and mud parameters in the target multi-source sensor data, online correction of the formation equivalent resistance torque model, which is used to estimate the torque required for the drill bit to cut under the current formation conditions; within each control cycle, based on the equipment capability boundary and the corrected formation equivalent resistance torque model, combined with the probability of anomaly occurrence and the type of anomaly symptom, calculating the upper limit value of torque changing over time, thereby generating the dynamic torque limit curve, and applying a preset slope constraint to the rate of change of the dynamic torque limit curve.

5. A dynamic torque limiting and early warning linkage edge control system for top drive systems, characterized in that, It includes a data acquisition module, a risk assessment module, a curve generation module, and a linkage control module; The data acquisition module is used to: acquire multi-source sensor data that reflects the real-time working conditions of drilling in real time, and perform unified semantic point table management to obtain target multi-source sensor data. The risk assessment module is used to: use a trained TCN-Transformer model to assess the risk of the time series composed of the target multi-source sensor data, and obtain the probability of anomaly occurrence and the type of anomaly symptoms; The curve generation module is used to: generate a dynamic torque limit curve based on the probability of the anomaly and the type of the anomaly, combined with the rated performance parameters of the top drive system and the target multi-source sensor data; The linkage control module is used to: when the probability of the anomaly exceeds a preset threshold, lower the current upper limit of the torque of the dynamic torque limit curve, reduce the upper limit of the rotational speed of the top drive system, and limit the drill string lifting speed or drilling pressure, while sending a coordinated control signal to the well control safety system to trigger the well control related devices to enter the prevention state in advance.

6. A dynamic torque limiting and early warning linkage edge control system for top drive systems according to claim 5, characterized in that, The target multi-source sensing data includes at least one of top drive torque, rotational speed, hook load, drilling pressure, mud pump pressure, and drill string vibration signal; the multi-source sensing data includes raw data frames collected by multiple sensors; the data acquisition module is specifically used to: aggregate the raw data frames to the edge computing unit via a distributed soft bus; perform protocol type identification and register address space parsing on the raw data frames to identify the protocol type corresponding to the raw data frames; and convert the observations corresponding to the target multi-source sensing data in the raw data frames into engineering data identified by standard semantic objects according to a pre-configured semantic-point mapping file, and use the engineering data as the target multi-source sensing data.

7. A dynamic torque limiting and early warning linkage edge control system for top drive systems according to claim 5, characterized in that, The risk assessment module is specifically used for: dividing the time series of the target multi-source sensor data into multiple time windows according to a preset sliding window length; extracting at least one feature from the target multi-source sensor data within each time window, including time-domain statistics, envelope, frequency-domain energy, spectral peak stability, and amplitude variation rate, to obtain a feature tensor; inputting the feature tensor into the trained TCN-Transformer model, where the TCN layer in the trained TCN-Transformer model captures local temporal features, and the Transformer layer in the trained TCN-Transformer model extracts global dependency features across channels and across windows; and outputting the probability of anomaly occurrence and the type of anomaly symptom based on the local temporal features and the global dependency features.

8. A dynamic torque limiting and early warning linkage edge control system for top drive systems according to claim 5, characterized in that, The curve generation module is specifically used for: obtaining the equipment capability boundary of the top drive system, which is determined by the rated torque, short-term torque, thermal limit value, and speed-torque envelope relationship of the top drive system and the rotation system; correcting the formation equivalent resistance torque model online based on the speed, drilling pressure, mud pump pressure, well depth, friction, and mud parameters in the target multi-source sensor data, which is used to estimate the torque required for the drill bit to cut under the current formation conditions; and calculating the upper limit value of torque changing with time based on the equipment capability boundary and the corrected formation equivalent resistance torque model, combined with the probability of anomaly occurrence and the type of anomaly symptom, thereby generating the dynamic torque limit curve, and applying a preset slope constraint to the rate of change of the dynamic torque limit curve.

9. An electronic device, characterized in that, The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the dynamic torque limiting and early warning linkage edge control method for a top drive system as described in any one of claims 1 to 4.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the dynamic torque limiting and early warning linkage edge control method for a top drive system as described in any one of claims 1 to 4.

Images