Core drill fault early warning method based on multi-step prediction
By generating virtual fault samples and using a CNN-LSTM-Attention model for multi-step prediction, the problem of lagging early warning of drilling rig faults in scenarios with few samples is solved, achieving early and accurate fault warning and adapting to complex drilling conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-10
- Publication Date
- 2026-04-07
AI Technical Summary
Existing drilling rig fault early warning methods suffer from delayed warnings and insufficient accuracy in scenarios with few samples, making it difficult to provide early and accurate fault warnings.
A multi-step prediction-based approach is adopted, which expands the training data by generating virtual fault samples, combines the CNN-LSTM-Attention model for multi-time step prediction, uses the Seq2Seq structure to achieve end-to-end fault warning, and sets dynamic safety thresholds for real-time comparison.
It significantly improves the accuracy and lead time of fault warnings in real-world scenarios with few samples, provides a sufficient window for emergency response, adapts to changes in rock strata, and achieves earlier and more accurate fault warnings.
Smart Images
Figure CN121808366A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of machine learning and fault prediction technology, and more specifically, to a fault early warning method for core drilling rigs based on multi-step prediction. Background Technology
[0002] As core equipment in geological exploration and resource development, the real-time monitoring and fault early warning of drilling rigs are crucial for ensuring project safety and improving drilling efficiency. Currently, drilling rig fault early warning mainly employs methods such as fixed threshold alarms, expert experience-based diagnostics, traditional machine learning-based fault diagnosis, and early warning based on single prediction models.
[0003] Alarm methods based on fixed thresholds achieve fault alarms by setting fixed threshold ranges for drilling rig parameters. While this method is simple in principle and low in implementation cost, it suffers from significant early warning lag, only issuing alarms after a fault occurs, failing to provide operators with sufficient reaction time. Furthermore, due to the complex and variable nature of drilling conditions, fixed thresholds are difficult to adapt to the actual needs of different formations and drilling tool combinations, easily leading to false alarms or missed alarms.
[0004] Diagnostic methods based on expert experience rely on the accumulated experience of on-site operators, judging faults by observing parameter change trends. Although this method is still used in actual engineering, it has inherent defects such as strong subjectivity, poor replicability, and difficulty in standardization, and cannot meet the needs of modern intelligent drilling for accurate early warning.
[0005] Traditional machine learning-based fault diagnosis methods employ algorithms such as Support Vector Machines and Random Forests to classify and identify operating conditions by learning from historical fault data, demonstrating a certain level of intelligent diagnostic capability compared to threshold-based alarm methods. However, these methods heavily rely on training with a large number of labeled fault samples, while actual drilling processes involve scarce fault data, resulting in a severe imbalance between positive and negative samples. Furthermore, traditional machine learning methods struggle to effectively capture the temporal dependencies between drilling rig parameters, exhibiting significantly insufficient early identification capability for progressive faults.
[0006] Early warning methods based on a single prediction model employ traditional time series analysis or a single deep learning model to predict future parameter trends by learning from historical data. However, traditional time series methods perform poorly when dealing with complex drilling rig systems characterized by multiple parameters, nonlinearity, and strong coupling; while a single LSTM model, although capable of capturing long-term dependencies, has limited ability to extract local features, making its accuracy in multivariate, multi-scale prediction tasks insufficient for engineering requirements. More importantly, most existing methods focus on single-step prediction and cannot provide continuous, multi-time-step early warning information.
[0007] How to achieve earlier and more accurate fault warnings in scenarios with few samples is a technical problem that urgently needs to be solved. Summary of the Invention
[0008] The purpose of this invention is to provide a multi-step prediction-based fault early warning method for core drilling rigs, which can improve the accuracy of fault early warning and extend the early warning time in real-world scenarios with few samples.
[0009] This invention provides a multi-step prediction-based method for early warning of core drilling rig failures, comprising the following steps: S1: Obtain the drilling parameters of the core drilling rig, preprocess the drilling parameters to obtain normal samples; S2: Generate virtual fault samples based on the normal samples to obtain training and test sets; S3: Train the early warning model using the training set to obtain a trained early warning model; use the trained early warning model to predict the test set to obtain a parameter prediction sequence.
[0010] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described multi-step prediction-based core drilling rig fault early warning method.
[0011] Implementing the multi-step prediction-based core drilling rig fault early warning method provided by this invention has the following beneficial effects: This invention addresses the industry challenges of scarce fault samples and severe imbalance between positive and negative samples in the drilling process. It proposes a virtual sample construction method based on the fusion of mechanism analysis and data perturbation. Without relying on a large amount of real fault data, this method combines the physical mechanism characteristics of typical drilling rig faults (such as drill bit burning, drill bit jamming, drill bit stuck, and drill bit breaking) and performs directional perturbation and superposition on parameter deviation values based on normal samples to generate physically meaningful virtual fault samples. This effectively expands the training dataset, breaks through the bottleneck of machine learning model training under few sample conditions, and significantly improves the model's generalization ability and robustness in real few sample scenarios. This invention constructs a Seq2Seq structure of "encoder-decoder" for multi-parameter time-series prediction design of drilling rigs. The encoder utilizes the local feature extraction capability of the convolutional neural network (CNN) module to capture parameter fluctuation features and linkage patterns within a local time window, and utilizes the long-term dependency modeling capability of the long short-term memory network (LSTM) module to learn long-term temporal dependencies. To address the problem of insufficient attention to key temporal information in traditional methods, the invention uses the key information focusing capability of the attention mechanism to adaptively assign differentiated weights to features at different time steps, highlighting key fault symptom information, improving the model's ability to identify early fault features, and enabling the model to more accurately capture the evolution law of progressive faults, achieving earlier and more accurate fault warnings. To address the issues of insufficient early warning timeliness and low accuracy across multiple time steps, this invention employs a Seq2Seq multi-time step prediction framework to achieve end-to-end prediction from a complete drilling cycle to future multiple time steps. It can output a sequence of parameter deviation values for a specified future step length (e.g., 12 seconds, 24 seconds), significantly improving the lead time and practicality of early warnings. Finally, based on historical normal data and combined with the 3σ principle, dynamic safety thresholds are set for each parameter for real-time comparison, effectively extending the early warning lead time and providing sufficient emergency response windows for on-site operators, thus achieving adaptive early warning based on changes in rock strata. Attached Figure Description
[0012] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a schematic diagram of the operation steps of the multi-step prediction-based core drilling rig fault early warning method provided by the present invention; Figure 2 This is a flowchart illustrating the multi-step prediction-based fault early warning method for core drilling rigs provided by the present invention. Figure 3 This is a schematic diagram of the core drilling rig data acquisition system provided by the present invention; Figure 4 This is a schematic diagram of the CNN-LSTM-Attention early warning model structure provided by the present invention; Figure 5 This is a schematic diagram of the input sample provided by the present invention; Figure 6 This is a schematic diagram of the LSTM timing modeling module structure provided by the present invention. Detailed Implementation
[0013] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0014] Figure 1A schematic diagram of the multi-step prediction-based core drilling rig fault early warning method of this embodiment is shown. In this embodiment, the multi-step prediction-based core drilling rig fault early warning method includes the following steps: S1: Obtain the drilling parameters of the core drilling rig, preprocess the drilling parameters to obtain normal samples.
[0015] In one exemplary embodiment, the drilling parameters include rotational speed, drilling speed, rod chamber pressure, rodless chamber pressure, and torque.
[0016] In one exemplary embodiment, the preprocessing includes smoothing, bias processing, removal of outlier data points, and normalization.
[0017] In one exemplary embodiment, the smoothing process includes: smoothing the original time-series data of drilling parameters using a moving average method to obtain smoothed data, as shown in the formula:
[0018]
[0019] in, The first time series data representing the drilling parameters n One data point; is the value of the nth data point after smoothing; m is the size of the sliding window.
[0020] In one exemplary embodiment, the deviation value processing includes: calculating the difference between parameters at adjacent time points, enhancing the model's ability to perceive dynamic changes in parameters, and converting the smoothed data into a parameter increment sequence that better reflects changes in operating conditions, as shown in the formula:
[0021] in, For parameter increments; It is the first i The actual values of drilling data at each time point include four characteristic parameters: rotational speed, torque, drilling speed, and axial pressure.
[0022] In one exemplary embodiment, the removal of outlier data points includes: based on mathematical statistics... The principle is to filter parameter deviation values and calculate the mean of the parameter increment sequence under normal operating conditions. with standard deviation Then remove those that fall into Abnormal data points outside the interval are converted into normal data sequences, which effectively remove outliers caused by accidental interference and ensure the stability of the subsequent model learning foundation.
[0023] In one exemplary embodiment, the normalization includes: linearly mapping the data in the normal data sequence to the [0,1] interval using a max-min normalization method to obtain normal samples, which are used to eliminate the differences in the units and numerical ranges between different drilling parameters, accelerate model convergence, and improve training stability, as shown in the formula:
[0024] in, This is the normalized data for normal samples; This is the data before normalization for normal samples; This is the minimum value for normal samples; This represents the maximum value for normal samples.
[0025] S2: Generate virtual fault samples based on the normal samples to obtain training and test sets.
[0026] In one exemplary embodiment, step S2 specifically includes: S21: Based on the physical manifestations of typical drilling rig faults, obtain the fault-parameter response relationship; S22: Based on the normal sample, the parameter deviation value is directionally disturbed and superimposed using the fault-parameter response relationship to obtain a virtual fault sample; S23: Mix the virtual fault samples with normal samples and divide them according to a preset ratio to obtain a training set and a test set.
[0027] In one exemplary embodiment, the preset ratio is 7:3.
[0028] S3: Train the early warning model using the training set to obtain a trained early warning model; use the trained early warning model to predict the test set to obtain a parameter prediction sequence.
[0029] In one exemplary embodiment, the warning model includes an encoder, an attention mechanism, a decoder, and an output layer; The encoder is used to extract deep features from the input sequence; The attention mechanism is used to adaptively assign differentiated weights to the hidden states at different time steps, thereby highlighting the key symptom information most relevant to fault warnings. The decoder is used to generate a prediction sequence with a specified future step length step by step based on the global context information provided by the encoder. The output layer is used to output a predicted sequence of 4D parameter deviation values for multiple future time steps.
[0030] In one exemplary embodiment, the encoder includes a cascaded CNN feature extraction module and an LSTM temporal modeling module; The CNN feature extraction module is used to capture the fluctuation features and linkage patterns of drilling parameters within a local time window using a one-dimensional convolutional layer. The LSTM time series modeling module is used to receive the high-order feature sequence output by the CNN feature extraction module, and use the gating mechanism to learn the long-term dependencies in the drilling parameter sequence in order to capture the evolution law of progressive faults.
[0031] In one exemplary embodiment, the CNN feature extraction module includes a first convolutional layer and a second convolutional layer; the first convolutional layer includes 16 one-dimensional convolutional kernels of size 3; the second convolutional layer includes 32 one-dimensional convolutional kernels of size 5; after convolution, both the first and second convolutional layers undergo ReLU activation and max pooling with a stride of 2 to progressively extract and compress features, and output a high-order feature sequence.
[0032] In one exemplary embodiment, the LSTM time-series modeling module includes a long short-term memory network.
[0033] In one exemplary embodiment, the Long Short-Term Memory (LSTM) network achieves the filtering and retention of long-term temporal information through the collaborative operation of input gates, forget gates, output gates, and candidate cell states.
[0034] In one exemplary embodiment, the attention mechanism is configured to: calculate the attention score at each time step, normalize the attention score using the Softmax function to obtain the attention weight, and finally obtain a context vector condensed with key information by weighted summation of all hidden states, as shown in the formula:
[0035]
[0036]
[0037] in, , and These represent attention score, attention weight, and context vector, respectively. This is the weight matrix. For bias terms, This represents the initial state of the decoder; The current hidden state is T; the length of the time series is T.
[0038] In one exemplary embodiment, the decoder includes a long short-term memory network, and the initial state of the decoder is initialized by the final state of the encoder and the attention context vector.
[0039] In one exemplary embodiment, the output layer adopts a fully connected network structure, and the output dimension of the output layer matches the target prediction time step.
[0040] In one exemplary embodiment, the early warning model is trained using the Adam optimizer with mean squared error as the loss function.
[0041] In one exemplary embodiment, the core drilling rig fault early warning method based on multi-step prediction further includes: performing core drilling rig fault early warning according to the parameter prediction sequence and dynamic safety threshold range.
[0042] In one exemplary embodiment, the step of providing early warning of core drilling rig failures based on the parameter prediction sequence and dynamic safety threshold range includes: Based on the historical normal drilling data in the normal sample, the following was adopted: The principle is to set a dynamic safety threshold range for the deviation values of each drilling parameter; If the parameter deviation value of the parameter prediction sequence exceeds the dynamic safety threshold range, an early warning is triggered.
[0043] In some embodiments, the above-described multi-step prediction-based core drilling rig fault early warning method can also be implemented in the following ways.
[0044] like Figure 2 The diagram shows a flowchart of a multi-step prediction-based core drilling rig fault early warning method. In this embodiment, the multi-step prediction-based core drilling rig fault early warning method includes the following steps: Step 1) Install the sensor system on the core drilling rig, including a speed sensor, a draw rope displacement sensor, a pressure sensor, and a wireless torque sensor to collect drilling parameters in real time. These parameters include speed, drilling rate, rod chamber pressure, rodless chamber pressure, and torque. Figure 3 The image shows the data acquisition system of a core drilling rig.
[0045] The collected raw data underwent the following preprocessing: (1) The moving average method is used to smooth the original time series data. For a set of time series data , , ..., If the sliding window size is set to m Then the first i The moving average of data points (when It can be calculated based on formulas (4-1) and (4-2).
[0046] (4-1) (4-2) (2) By using deviation value processing to enhance the model's ability to perceive dynamic changes in parameters, the original absolute parameter values are converted into a parameter increment sequence that better reflects changes in operating conditions. According to (4-3), this increment is obtained by calculating the difference between parameters at adjacent time points. Wherein, It is the first i The actual values of drilling data at each time point include four characteristic parameters: rotational speed, torque, drilling speed, and axial pressure.
[0047] (4-3) (3) A reliable normal operating condition sample library is constructed using safety threshold processing, based on mathematical statistics. The principle is to screen the parameter deviation values and calculate the mean value of the parameter deviation values under normal operating conditions. ) and standard deviation ( ), and then remove those that fall into Outlier data points outside the specified interval. This process effectively removes outliers caused by accidental interference, ensuring the stability of the subsequent model learning foundation.
[0048] (4) Normalize the data to eliminate the differences in dimensions and numerical ranges between different drilling parameters, accelerate model convergence and improve training stability. Use the maximum-minimum normalization method to linearly map the data processed above to the [0,1] interval. (4-4) is its calculation formula.
[0049] (4-4) Step 2) Construction of Virtual Fault Samples Based on Mechanism Analysis. Based on the physical manifestations of typical drilling rig faults (such as drill bit burning, drill bit jamming, drill bit stuck, and drill bit breakage), the corresponding parameter change patterns are determined. For example, a stuck drill bit condition is characterized by a sharp increase in torque and a significant decrease in rotational speed and drilling speed. Based on the obtained normal samples, the parameter deviation values are directionally perturbed and superimposed according to the fault-parameter response relationship to generate physically meaningful virtual fault samples. The generated virtual fault samples are mixed with normal samples in a proportional ratio to construct a dataset for model training, where the ratio of the training set to the test set is 7:3.
[0050] Step 3) Construct a CNN-LSTM-Attention early warning model, employing an encoder-decoder Seq2Seq structure specifically designed for multi-parameter time-series prediction of drilling rigs. This model uses a complete drilling cycle as the input time window, inputting a preprocessed 4D parameter deviation value sequence into the model and directly outputting the parameter prediction sequence for multiple future time steps, thus achieving end-to-end fault early warning. The core architecture of the model consists of an encoder, an attention mechanism, and a decoder; its structure diagram is shown in the reference diagram. Figure 4 , Figure 5 This is the input sample.
[0051] The encoder, responsible for deep feature extraction from the input sequence, consists of a CNN module and an LSTM module connected in series. First, the CNN feature extraction module uses one-dimensional convolutional layers to capture the fluctuation features and linkage patterns of drilling parameters within a local time window. This module employs a two-layer convolutional structure: the first layer uses 16 one-dimensional convolutional kernels of size 3, and the second layer uses 32 one-dimensional convolutional kernels of size 5. After each convolution, a ReLU activation function is applied followed by max pooling with a stride of 2 to progressively extract and compress features, outputting a high-order feature sequence. Subsequently, the LSTM temporal modeling module receives the feature sequence output from the CNN module and uses its gating mechanism to learn the long-term dependencies in the drilling parameter sequence to capture the evolutionary patterns of progressive faults, referencing... Figure 6 Long Short-Term Memory (LSTM) networks utilize the collaborative operations of input gates, forget gates, output gates, and candidate cell states to filter and retain long-term temporal information. Equation (4-5) represents the input gate. Calculation, (4-6) is the forgetting gate Calculation, (4-7) is the output gate Calculations (4-8) represent the candidate cell states. Calculate (4-9) for the current cell state. Calculate (4-10) as the current hidden state. Calculation, where The input features are for time step t. This is the hidden state from the previous moment. and Here are the weight matrix and bias vector for the corresponding gate. for Activation function It is the hyperbolic tangent function. express product, The module ultimately outputs the hidden state sequence for all time steps. .
[0052] (4-5) (4-6) (4-7) (4-8) (4-9) (4-10) An attention mechanism is introduced between the encoder and decoder to adaptively assign differentiated weights to the hidden states at different time steps, thereby highlighting the key symptom information most relevant to fault warnings. First, the attention score for each time step is calculated, as shown in Equation (4-11), where... This is the weight matrix. For bias terms, This represents the initial state of the decoder. Subsequently, the score is normalized using the Softmax function to obtain the attention weights, as shown in Equation (4-12). Finally, a context vector condensing key information is obtained by weighted summation of all hidden states, as shown in Equation (4-13).
[0053] (4-11) (4-12) (4-13) The decoder is also composed of LSTM units, and its initial state is initialized by the encoder's final state and the attention context vector. Based on the global context information provided by the encoder, the decoder gradually generates prediction sequences for future specified steps.
[0054] The output layer adopts a fully connected network structure, and its output dimension matches the target prediction time step. For example, for a 12-second warning requirement, the output dimension is set to 12 to correspond to 12 time steps. A linear activation function is used to directly output the 4-dimensional parameter deviation value prediction sequence for future multiple time steps.
[0055] The model is trained using the Adam optimizer with mean squared error (MSE) as the loss function. The Adam optimizer dynamically adjusts the learning rate of each parameter by calculating the first-moment and second-moment estimates of the parameters, and its update rules are shown in formulas (4-14) to (4-17). Let be the gradient at time t. and The exponential decay rate is estimated by moments. For learning rate, These are constants added to maintain numerical stability.
[0056] (4-14) (4-15) (4-16) (4-17) Step 4) Based on historical normal drilling data, adopt... The principle is to set dynamic safety threshold ranges for the deviation values of each drilling parameter (rotation speed, torque, drilling speed, axial pressure), and the threshold range is adaptively adjusted according to changes in the rock strata. The prediction results of multiple time steps (parameter sequences for the next 12 seconds and 24 seconds) are compared with the dynamic safety thresholds in real time. If the parameter deviation values of multiple consecutive time steps in the prediction sequence exceed the threshold range, an early warning is triggered.
[0057] This embodiment provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described multi-step prediction-based core drilling rig fault early warning method.
[0058] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.
Claims
1. A core drilling rig fault early warning method based on multi-step prediction, characterized in that, Includes the following steps: S1: Obtain the drilling parameters of the core drilling rig, preprocess the drilling parameters to obtain normal samples; S2: Generate virtual fault samples based on the normal samples to obtain training and test sets; S3: Train the early warning model using the training set to obtain a trained early warning model; use the trained early warning model to predict the test set to obtain a parameter prediction sequence.
2. The core drilling rig fault early warning method based on multi-step prediction according to claim 1, characterized in that, Also includes: Core drilling rig fault warnings are generated based on the predicted sequence of parameters and the dynamic safety threshold range.
3. The core drilling rig fault early warning method based on multi-step prediction according to claim 2, characterized in that, The method of providing early warning of core drilling rig failures based on the predicted sequence of parameters and the dynamic safety threshold range includes: Based on the historical normal drilling data in the normal sample, the following was adopted: The principle is to set a dynamic safety threshold range for the deviation values of each drilling parameter; If the parameter deviation value of the parameter prediction sequence exceeds the dynamic safety threshold range, an early warning is triggered.
4. The core drilling rig fault early warning method based on multi-step prediction according to claim 1, characterized in that, The drilling parameters include rotational speed, drilling speed, rod chamber pressure, rodless chamber pressure, and torque.
5. The core drilling rig fault early warning method based on multi-step prediction according to claim 1, characterized in that, The preprocessing includes smoothing, deviation value processing, removal of outlier data points, and normalization.
6. The core drilling rig fault early warning method based on multi-step prediction according to claim 1, characterized in that, Step S2 specifically includes: S21: Based on the physical manifestations of typical drilling rig faults, obtain the fault-parameter response relationship; S22: Based on the normal sample, the parameter deviation value is directionally disturbed and superimposed using the fault-parameter response relationship to obtain a virtual fault sample; S23: Mix the virtual fault samples with normal samples and divide them according to a preset ratio to obtain a training set and a test set.
7. The core drilling rig fault early warning method based on multi-step prediction according to claim 1, characterized in that, The early warning model includes an encoder, an attention mechanism, a decoder, and an output layer; the encoder is used to extract deep features from the input sequence; the attention mechanism is used to adaptively assign differentiated weights to the hidden states at different time steps, thereby highlighting the key symptom information most relevant to the fault warning; the decoder is used to gradually generate a prediction sequence for a specified future step based on the global context information provided by the encoder. The output layer is used to output a predicted sequence of 4D parameter deviation values for multiple future time steps.
8. The core drilling rig fault early warning method based on multi-step prediction according to claim 7, characterized in that, The encoder includes a cascaded CNN feature extraction module and an LSTM temporal modeling module. The CNN feature extraction module is used to capture the fluctuation features and linkage patterns of drilling parameters within a local time window using a one-dimensional convolutional layer. The LSTM temporal modeling module is used to receive the high-order feature sequence output by the CNN feature extraction module and use a gating mechanism to learn the long-term dependencies in the drilling parameter sequence in order to capture the evolution law of progressive faults.
9. The core drilling rig fault early warning method based on multi-step prediction according to claim 1, characterized in that, The early warning model is trained using the Adam optimizer, with mean squared error as the loss function.
10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the multi-step prediction-based core drilling rig fault early warning method as described in any one of claims 1-9.