CN-121997089-A - Drilling machine fault diagnosis, alarm and processing method and device

CN121997089ACN 121997089 ACN121997089 ACN 121997089ACN-121997089-A

Abstract

The application provides a drilling machine fault diagnosis, alarm and processing method and device, wherein the method comprises the steps of obtaining operation data and logging data of a target drilling machine, wherein the operation data comprise at least one of motor torque, motor rotation speed, motor power and motor temperature, the logging data comprise at least one of drilling pressure, rotation speed, pumping pressure, well depth and torque, performing time stamp alignment, data normalization, feature cleaning and missing compensation on the operation data and the logging data to construct a unified time sequence multi-modal feature set, inputting the unified time sequence multi-modal feature set into a pre-trained large language model, performing abnormal diagnosis on the target drilling machine, and outputting an improvement suggestion. The technical problems of low fault diagnosis efficiency and low accuracy of the existing drilling machine are solved through the scheme, and the technical effects of improving the accuracy, generalization and response speed of fault diagnosis of the drilling machine are achieved.

Inventors

YAN ZHI
SONG XIANZHI
ZHU ZHAOPENG
LI GENSHENG
TIAN SHOUCENG
HUANG ZHONGWEI
SHI HUAIZHONG
ZHOU MENGMENG

Assignees

中国石油大学(北京)

Dates

Publication Date: 20260508
Application Date: 20260408

Claims (10)

1. The fault diagnosis, alarm and processing method for the drilling machine is characterized by comprising the following steps of: Acquiring operation data and logging data of a target drilling machine, wherein the operation data comprise at least one of motor torque, motor rotation speed, motor power and motor temperature, and the logging data comprise at least one of weight on bit, rotation speed, pump pressure, well depth and torque; Performing time stamp alignment, data normalization, feature cleaning and missing compensation on the operation data and the logging data to construct a unified time sequence multi-mode feature set; and inputting the unified time sequence multi-modal feature set into a pre-trained large language model, performing anomaly diagnosis on the target drilling machine, and outputting improvement suggestions.
2. The method of claim 1, wherein inputting the unified time series multimodal feature set into a pre-trained large language model, diagnosing anomalies on the target rig, and outputting improvement suggestions, comprises: Carrying out multi-layer full-connection transformation and nonlinear mapping on the numerical feature vector to obtain numerical feature embedding; the text sequence is embedded with codes through word embedding and position coding; Mapping the abnormal weight output by the abnormal clustering module into a group of abnormal indication embedments; Embedding the numerical characteristics, the coding and the abnormal indication to form a unified input sequence so as to realize cross-modal fusion of the numerical characteristics, the text characteristics and the abnormal weights in the same hidden space through a self-attention mechanism; and processing the unified input sequence through an encoder in the large language model, obtaining a hidden state sequence through multi-head self-attention calculation, and outputting an abnormality diagnosis result and an improvement suggestion.
3. The method of claim 2, wherein processing the unified input sequence through an encoder in the large language model, obtaining a hidden state sequence through multi-headed self-attention calculation, and outputting an abnormality diagnosis result and an improvement suggestion, comprises: And performing task decoupling through a plurality of task heads arranged at the top of the output of the encoder to obtain a state prediction result and an improvement suggestion.
4. The method of claim 3, wherein the plurality of task heads comprises a status determination head, a reason trace generation head, and a process suggestion generation head, wherein: the state judgment head is used for carrying out linear mapping and softmax on the window-level hidden state sequence to obtain a state prediction result; The reason tracing generation head is used for generating a reason description sequence based on the autoregressive hidden state sequence at the output position of the causal chain; The processing suggestion generation head is used for generating a suggestion sequence in the processing suggestion output position based on the hidden state sequence, the generated state label and the reason description in an autoregressive mode.
5. The method of claim 1, wherein the large language model is trained according to a loss function as follows: ; Wherein, the As a value of the total loss, In order to determine the loss in the status, For the reason of the loss of the source tracing language modeling, In order to suggest the generation of a penalty, In order to pay attention to the loss of the significant constraint, 、、、 The weight coefficient for each loss term.
6. The method according to claim 5, wherein: The state judgment loss is expressed as: ; Wherein, the The loss value is judged for the state, As a true state label, For the model prediction result, k is a sample index for state judgment, and Σ k represents that the logarithmic product term of the real state labels of all samples and the model prediction result is accumulated; the source tracing language modeling loss is expressed as: ; Wherein, the The loss value is modeled for the source tracing language, Describing the token for the manually noted or sorted standard reason, t representing the time step of the sequence generation, Representing hidden state sequence, p # ) Representing the conditional probability distribution given by the model; The advice generation penalty is expressed as: ; ; Wherein, the In order to suggest the generation of a loss value, In order to order the loss of the proposed ordering for the plurality of candidates, The scores of the different candidate suggestions are respectively, Representing the standard advice sequence at the first The actual target token for the time step, Representing a set of sample pairs that should be higher; the attention deficit constraint loss is expressed as: ; Wherein, the In order to pay attention to the significance constraint loss value, In order to trade-off the coefficients, For the normalized anomaly weight to be the same, For the abnormal weight to be given, For the concentration distribution of the layer, And the position set corresponds to the abnormal point.
7. The method of claim 1, wherein inputting the unified time series multimodal feature set into a pre-trained large language model for anomaly diagnosis of the target drilling rig comprises: the large language model automatically learns compact areas distributed by normal working conditions, and identifies and divides outliers and density clusters thereof into abnormal state areas; And visually presenting the clustering division result in a two-dimensional map mode, wherein a first color region represents a high-density abnormal cluster, and a second color region represents a low-risk or normal state.
8. The method of claim 7, wherein the large language model automatically learns compact regions distributed by normal conditions and classifies outlier and its density cluster identification into abnormal state regions, comprising: Constructing a high-dimensional feature vector by taking a time window as granularity, and extracting the following statistics and morphological features for a sliding time window with the length of T: ; Wherein, the In order to make statistics of the morphology features obtained, Is the mean value of the two values, Is the standard deviation of the two-dimensional image, At the point of maximum value of the energy, As a correlation coefficient, P is a pressure parameter during drilling by a drilling machine, and WOB is downward pressure applied to a drill bit; Constructing a sample set for all window features And performing standardization to obtain a normalized feature matrix : ; Wherein, the For the normalized feature vector corresponding to the i-th sample, i is the index of the sample, To estimate the resulting mean value over the training set, To estimate the resulting standard deviation over the training set, N represents the number of samples and the number of sliding time windows, and d represents the dimension of each window feature vector; Density clustering is carried out on the normalized feature matrix, the neighborhood radius and the minimum neighbor number are defined, and any sample is subjected to Calculating the neighborhood: ; Wherein, the Representing a sample A kind of electronic device A neighborhood region is defined in the region of the first region, Representing a neighborhood radius threshold value, Representing a normalized feature vector corresponding to the jth sample; At the position of In the case of (a), will Marking as core point, forming cluster label according to density reachable rule Points that fail to fall into any cluster are labeled noise point-1, where, Is the minimum number of neighbors.
9. A device for diagnosing, alarming and processing faults of a drilling machine, comprising: the acquisition module is used for acquiring operation data and logging data of the target drilling machine, wherein the operation data comprise at least one of motor torque, motor rotating speed, motor power and motor temperature, and the logging data comprise at least one of weight on bit, rotating speed, pump pressure, well depth and torque; The construction module is used for performing time stamp alignment, data normalization, feature cleaning and missing compensation on the operation data and the logging data so as to construct a unified time sequence multi-mode feature set; And the diagnosis module is used for inputting the unified time sequence multi-modal feature set into a pre-trained large language model, carrying out abnormality diagnosis on the target drilling machine and outputting improvement suggestions.
10. An electronic device comprising a processor and a memory for storing processor-executable instructions, wherein the processor, when executing the instructions, performs the steps of the method of any one of claims 1 to 8.

Description

Drilling machine fault diagnosis, alarm and processing method and device Technical Field The application belongs to the technical field of oil and gas drilling, and particularly relates to a method and a device for fault diagnosis, alarm and treatment of a drilling machine. Background In the oil and gas drilling process, the drilling machine is used as core mechanical equipment, and the running state of the drilling machine has decisive effect on the safety, efficiency and cost control of well site operation. The existing fault detection of the drilling machine mostly depends on manual experience and preset rules, such as monitoring the ultra-limit value of a specific sensor or performing periodic manual inspection. The method has obvious defects when facing complex geological conditions, high-frequency working condition switching or underground fault mode evolution, and is easy to cause false alarm and missing alarm and even serious fault shutdown of the drilling machine. Along with popularization of intelligent sensors and improvement of logging real-time data acquisition capability, drilling site data sources are more and more, but multisource data are often scattered and independent, and a unified semantic representation and a systematic reasoning mechanism are lacked, so that complex state diagnosis cannot be effectively supported. The existing machine learning model has complicated training flow and poor adaptability to environmental changes, is difficult to update and upgrade once deployed, and mostly cannot provide clear diagnosis explanation and processing suggestions, so that decision support is difficult to provide for first-line operators. No effective solution has been proposed at present for how to accurately diagnose the drill failure. Disclosure of Invention The application aims to provide a method and a device for diagnosing, alarming and processing faults of a drilling machine, which can improve the accuracy, generalization and response speed of fault diagnosis of the drilling machine. The application provides a fault diagnosis, alarm and processing method and device for a drilling machine, which are realized as follows: a fault diagnosis, alarm and processing method for a drilling machine comprises the following steps: Acquiring operation data and logging data of a target drilling machine, wherein the operation data comprise at least one of motor torque, motor rotation speed, motor power and motor temperature, and the logging data comprise at least one of weight on bit, rotation speed, pump pressure, well depth and torque; Performing time stamp alignment, data normalization, feature cleaning and missing compensation on the operation data and the logging data to construct a unified time sequence multi-mode feature set; and inputting the unified time sequence multi-modal feature set into a pre-trained large language model, performing anomaly diagnosis on the target drilling machine, and outputting improvement suggestions. In one embodiment, inputting the unified time series multimodal feature set into a pre-trained large language model, performing anomaly diagnosis on the target drilling machine, and outputting improvement suggestions, comprising: Carrying out multi-layer full-connection transformation and nonlinear mapping on the numerical feature vector to obtain numerical feature embedding; the text sequence is embedded with codes through word embedding and position coding; Mapping the abnormal weight output by the abnormal clustering module into a group of abnormal indication embedments; Embedding the numerical characteristics, the coding and the abnormal indication to form a unified input sequence so as to realize cross-modal fusion of the numerical characteristics, the text characteristics and the abnormal weights in the same hidden space through a self-attention mechanism; and processing the unified input sequence through an encoder in the large language model, obtaining a hidden state sequence through multi-head self-attention calculation, and outputting an abnormality diagnosis result and an improvement suggestion. In one embodiment, the processing the unified input sequence by the encoder in the large language model, obtaining a hidden state sequence through multi-head self-attention calculation, and outputting an abnormality diagnosis result and an improvement suggestion includes: And performing task decoupling through a plurality of task heads arranged at the top of the output of the encoder to obtain a state prediction result and an improvement suggestion. In one embodiment, the plurality of task heads includes a status determination head, a reason trace generation head, and a process suggestion generation head, wherein: the state judgment head is used for carrying out linear mapping and softmax on the window-level hidden state sequence to obtain a state prediction result; The reason tracing generation head is used for generating a reason description sequence based on the autor