An ai-based blowout preventer product performance monitoring system and method

Abstract

The present application relates to the technical field of blowout preventer monitoring, and particularly relates to a blowout preventer product performance monitoring system and method based on AI, which comprises: a data acquisition module, which is used for acquiring blowout preventer operation data and outputting standardized multi-dimensional time sequence data flow; a feature extraction module, which is used for performing feature extraction on the standardized multi-dimensional time sequence data flow and outputting a multi-dimensional feature vector matrix; a state recognition module, which is used for constructing a state recognition model to recognize the current performance state of the blowout preventer according to the multi-dimensional feature vector matrix and generating a performance state probability distribution; a trend prediction module, which is used for constructing a trend prediction model according to the performance state probability distribution and a historical state sequence, predicting a future performance degradation trajectory and a remaining service life of the blowout preventer, and generating a prediction result; a decision warning module, which is used for constructing a multi-level warning system according to the prediction result and generating a maintenance decision scheme; and a learning optimization module, which is used for continuously optimizing the AI model performance according to the decision scheme and actual maintenance result feedback.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of blowout preventer monitoring, and particularly relates to a blowout preventer product performance monitoring system and method based on AI. BACKGROUND

[0002] The traditional blowout preventer performance monitoring scheme relies on manual inspection and single parameter monitoring in the data acquisition stage. Usually, only key pressure values are recorded by simple equipment, the sampling frequency is low, and multi-dimensional data such as vibration and temperature are lacking. Data processing relies on manual input into forms, which is prone to errors and omissions, and cannot form standardized time series data. This method cannot capture transient fault signals such as pressure fluctuations caused by sudden wear of sealing elements, and often the fault is not discovered until after the fault occurs.

[0003] In the state identification and early warning link, the traditional scheme relies on fixed threshold judgment. The operating personnel set the safety range of pressure and temperature parameters according to industry standards, and an alarm is triggered when the range is exceeded. However, the performance degradation of the blowout preventer is a gradual process, and the fixed threshold cannot identify early minor abnormalities such as vibration spectrum shift in the early stage of gate wear. At the same time, the alarm lacks a grading mechanism, which leads to slow response of the operating personnel to key early warnings and increases the risk of failure.

[0004] In terms of trend prediction and maintenance decision, the traditional scheme relies on experience-based judgment. Engineers develop fixed-cycle preventive maintenance plans based on equipment service life and historical failure records, such as mandatory maintenance every half year. This one-size-fits-all approach often leads to over-maintenance or insufficient maintenance, increasing costs and failing to accurately address individual device performance differences.

[0005] The traditional scheme lacks a dynamic optimization mechanism. Once the model and decision logic are determined, they remain unchanged for a long time and cannot adapt to factors such as blowout preventer aging and working condition changes. For example, a new device and a device used for five years have different performance implications under the same parameters, but the traditional scheme cannot dynamically adjust the judgment standard. At the same time, the deviation between the maintenance results and the predictions is not used to improve the system, resulting in difficulty in improving the monitoring accuracy and remaining at the level of passive response rather than active prediction. SUMMARY

[0006] To solve the above technical problems, the technical scheme adopted by the present application is: an AI-based blowout preventer product performance monitoring system, characterized by comprising:

[0007] A data acquisition module for acquiring blowout preventer operation data and outputting standardized multi-dimensional time series data streams;

[0008] A feature extraction module for extracting features from the standardized multi-dimensional time series data streams and outputting a multi-dimensional feature vector matrix;

[0009] A state recognition module is configured to construct a state recognition model to recognize a current performance state of the blowout preventer according to the multi-dimensional feature vector matrix, and generate a probability distribution of the performance state.

[0010] A trend prediction module is configured to construct a trend prediction model to predict a future performance degradation trajectory and a remaining service life of the blowout preventer according to the performance state probability distribution and a historical state sequence, and generate a prediction result.

[0011] A decision warning module is configured to construct a multi-level warning system according to the prediction result, and generate a maintenance decision scheme.

[0012] A learning optimization module is configured to continuously optimize AI model performance according to the decision scheme and actual maintenance result feedback.

[0013] The blowout preventer operation data includes: pressure sensor collected hydraulic system pressure data; vibration sensor monitored mechanical component vibration data; temperature sensor monitored key part temperature change data; displacement sensor detected ring blowout preventer action displacement data; flow sensor monitored hydraulic oil flow change data.

[0014] The feature extraction includes: hydraulic system pressure data extraction dynamic pressure mean and variance, pressure mutation index, FFT main frequency energy proportion, harmonic distortion rate; vibration data extraction resonance frequency band energy, fault feature frequency, wavelet energy entropy; temperature change data extraction temperature change rate, cumulative thermal effect, moving average residual; displacement data extraction action speed, acceleration, displacement curve fitting degree, displacement fluctuation variance; flow change data collection flow fluctuation coefficient, mutation jump point.

[0015] The multi-dimensional feature vector matrix is generated by: after standardizing the extracted features, a multi-dimensional feature vector matrix is constructed, each row corresponds to a time window, and includes multi-dimensional pressure features, multi-dimensional vibration features, multi-dimensional temperature features, multi-dimensional displacement features, and multi-dimensional flow features.

[0016] The state recognition model is constructed in the following manner: a CNN-LSTM hybrid neural network architecture is constructed, wherein the CNN is responsible for extracting the spatial mode of the multi-dimensional features, the LSTM is responsible for capturing the time sequence dependency, and the two are fused to form a comprehensive feature representation; the blowout preventer state is divided into four levels of normal, slight abnormality, moderate abnormality and severe abnormality according to industry standards and historical fault cases; a weighted cross-entropy loss function is used for dynamic sample weighting; historical operation data of the blowout preventer is collected to train and fine-tune the model, and a classification decision vector corresponding to the state is obtained.

[0017] The probability distribution of the performance state is generated in the following manner: the four original score values of the normal state, slight abnormality, moderate abnormality and severe abnormality obtained by the model are converted into a probability distribution by a Softmax function, which can be expressed as: ; wherein P(y=k|z) is the probability of the device being in state k given the input features z, z is the feature vector, w k is the classification decision vector corresponding to state k, is the degree of matching of the feature and the state decision hyperplane;

[0018] The probability distribution is temperature scaled and uncertainty quantified to obtain the probability distribution of the final performance state.

[0019] The way of predicting the future performance degradation trajectory and the remaining useful life of the blowout preventer is: a trend prediction model is constructed, historical health degree sequences are input, and future health degree curves are output; on the basis of trajectory prediction, a deep residual network is used to learn the mapping relationship between features and RUL; the real RUL of each time point is labeled using the full life cycle data of the blowout preventer, and the model is optimized in terms of mean square error; the future health degree degradation trajectory and the RUL estimation value are output.

[0020] The way of constructing a multi-level early warning system is: based on RUL and health degree threshold, 3-level early warning is divided, when RUL is greater than m days and health degree is greater than x%, it is level 1 early warning, the monitoring frequency is maintained, and the state report is generated daily; when RUL is greater than n days and less than m days, and health degree is greater than y% and less than x%, it is level 2 early warning, spare parts should be coordinated, downtime should be arranged, and maintenance procedures should be developed; when RUL is less than n days and health degree is less than y%, a standby blowout preventer should be started immediately, drilling operations should be stopped, and a maintenance team should be dispatched synchronously.

[0021] The way of continuously optimizing the performance of the AI model is: if the prediction deviation of a certain type of fault is large, it is analyzed whether key features are missed, effective features are selected using feature importance evaluation; feedback data is used for incremental learning to avoid full-volume retraining; for distribution offset data, domain adaptation algorithm is used to reduce distribution difference; if the maintenance cost of level 2 early warning is too high, the decision rule is adjusted, and the decision strategy is optimized through reinforcement learning.

[0022] An AI-based blowout preventer product performance monitoring method, characterized in that it comprises:

[0023] S1, collecting blowout preventer operation data and outputting standardized multi-dimensional time series data stream;

[0024] S2, performing feature extraction on the standardized multi-dimensional time series data stream and outputting a multi-dimensional feature vector matrix;

[0025] S3, constructing a state recognition model according to the multi-dimensional feature vector matrix to recognize the current performance state of the blowout preventer and generate a probability distribution of the performance state;

[0026] S4, according to the performance state probability distribution and the historical state sequence, a trend prediction model is constructed, a future performance degradation trajectory and a remaining service life of the blowout preventer are predicted, and a prediction result is generated;

[0027] S5, a multi-level early warning system is constructed according to the prediction result, and a maintenance decision scheme is generated;

[0028] S6, the AI model performance is continuously optimized according to the decision scheme and the actual maintenance result feedback.

[0029] In combination with all the technical solutions described above, the positive effects possessed by the present application are as follows: 1, the present application breaks through the information island limitation of traditional monitoring through full-dimensional data acquisition and standardization processing. Through the integration of multiple sensors such as pressure, vibration and temperature, the time series data stream is formed after noise reduction and normalization processing, and the holographic perception of the blowout preventer operating state is realized. Compared with the single-point sampling of traditional manual inspection, this method can capture transient anomalies and improve data integrity, laying a high-quality foundation for subsequent AI analysis and solving the problem of missed judgment caused by one-sided data of traditional solutions from the source.

[0030] 2, in the aspects of state recognition and trend prediction, the present application realizes the leap from passive alarm to active prediction. The performance state probability distribution is output by the state recognition model, and the slight degradation and normal state are accurately distinguished; the degradation trajectory and the remaining life are predicted in combination with the historical sequence, and the prediction error is reduced compared with the traditional experience estimation. This early warning capability can avoid the blowout risk caused by sudden failure, and provide accurate time basis for planned shutdown, significantly improving the safety of drilling operations.

[0031] 3, the present application greatly optimizes the operation and maintenance efficiency through the multi-level early warning and intelligent decision mechanism. According to the prediction result, a hierarchical response is triggered, forming a closed-loop management from strengthening monitoring to emergency shutdown, reducing the false alarm storm of traditional fixed threshold alarm; combined with the generated maintenance scheme under the working condition, the balance between spare parts inventory and drilling progress can be achieved, and the waste of resources caused by excessive maintenance can be avoided. In actual application, the non-planned downtime can be shortened, the annual maintenance cost can be reduced, and safety and economy can be considered.

[0032] 4, the present application guarantees the long-term effectiveness of the system through the continuous learning optimization mechanism. The model is dynamically updated through the maintenance result feedback, which adapts to the complex scenarios of blowout preventer aging and working condition changes, and solves the pain point of poor adaptability caused by parameter solidification in traditional solutions. With the accumulation of data, the state recognition accuracy can be gradually improved, the remaining life prediction accuracy can be continuously optimized, and a positive cycle of data, model and decision is formed, providing intelligent support for the whole life cycle management of the blowout preventer. BRIEF DESCRIPTION OF DRAWINGS

[0033] The application is further illustrated by the accompanying drawings, but the embodiments in the drawings do not constitute any limitation to the application, and other embodiments can be obtained by those skilled in the art without creative labor on the basis of the following drawings.

[0034] Figure 1 The system framework diagram of the system of the application. DETAILED DESCRIPTION

[0035] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0036] Referring to Figure 1 As shown in the drawings, the application provides an AI-based blowout preventer product performance monitoring system, which comprises a data acquisition module, a feature extraction module, a state recognition module, a trend prediction module, a decision warning module and a learning optimization module.

[0037] In a more specific application of the application, in the data acquisition module, the pressure sensor acquires hydraulic system pressure data by installing high-precision pressure sensors at key nodes of the blowout preventer hydraulic control circuit, including the inlet of the ram cylinder, the annular blowout preventer control pipeline and the accumulator outlet position. The sensor is connected to the hydraulic pipeline through threads, and real-time sensing of the dynamic pressure change of the hydraulic oil is performed at a sampling frequency of 10 Hz. The acquired original data is converted into a 4-20 mA current signal by a signal conditioning module, transmitted to an edge computing gateway, and then the peak noise generated by hydraulic impact is removed through digital filtering. Finally, the time-pressure form is stored in a time series database to form a continuous monitoring sequence of the hydraulic system pressure.

[0038] The vibration sensor monitors mechanical component vibration data. A piezoelectric acceleration sensor is selected and fixed on the surface of key mechanical components of the blowout preventer shell, the ram drive shaft and the annular rubber core gland through magnetic attraction or bolts. The sensor acquires vibration acceleration signals in three-dimensional directions X / Y / Z axes, and the sampling frequency reaches 1 kHz to capture high-frequency impact signals. After the original vibration signal is converted into a voltage signal by a charge amplifier, it is uploaded in real time through an industrial Ethernet, and the vibration interference of the drilling platform environment is eliminated through wavelet threshold denoising processing, and the mechanical fault characteristic frequency components are retained to provide high signal-to-noise ratio data for subsequent feature extraction.

[0039] The temperature sensor monitors the temperature change data of the key parts. The armored thermocouple and infrared temperature sensor are combined for monitoring. The thermocouple is embedded in the contact type measuring point of the sealing surface of the gate and the inner wall of the hydraulic oil pipeline, and transmits the temperature signal through the compensation lead. The infrared sensor is installed at a distance of 20-30 cm from the annular rubber core, and monitors the surface temperature of the rubber core in a non-contact manner. The sampling frequency is set to 1 Hz, and the data is time-aligned with the pressure and vibration data after cold-end compensation and linearization. When the temperature change rate exceeds 0.5°C / min, the sampling frequency is automatically increased to 10 Hz to focus on capturing the abnormal temperature rise process.

[0040] The displacement sensor detects the displacement data of the gate and the annular blowout preventer. A magnetostrictive displacement sensor is installed at the end of the piston rod of the gate oil cylinder, which moves synchronously with the piston rod through the built-in magnetic ring, and outputs the absolute displacement value of the gate opening and closing stroke in real time. The annular blowout preventer uses a pull wire type displacement sensor, with the pull wire end connected to the rubber core piston. The displacement sensor outputs a 4-20 mA analog signal, which is converted to digital after A / D conversion, and is checked for effectiveness in combination with the blowout preventer action command to eliminate abnormal jump values caused by mechanical jamming. Finally, a displacement-time curve is formed to intuitively reflect the action to the accuracy.

[0041] The flow sensor monitors the hydraulic oil flow change data. Turbine flow meters are installed at the outlet of the hydraulic pump and the total oil inlet pipeline of the blowout preventer. The turbine blade rotates to cut the magnetic lines to generate a pulse signal, and the pulse frequency is proportional to the flow. The pulse signal output by the sensor is converted to an instantaneous flow value by a counter, with a sampling frequency of 5 Hz and a total flow in a unit time. By comparing the flow difference between the inlet and outlet pipelines, it can be determined whether there is an internal leakage in the hydraulic system, and the fault positioning accuracy is improved by correlating the pressure data.

[0042] In the feature extraction module, the standardized multi-dimensional time series data stream is extracted for feature extraction, and a multi-dimensional feature vector matrix is output. The hydraulic system pressure data can extract dynamic pressure mean and variance, pressure mutation index, FFT main frequency energy proportion, and harmonic distortion rate. The vibration data can extract resonance band energy, fault characteristic frequency, and wavelet energy entropy. The temperature change data can extract temperature change rate, cumulative heat effect, and moving average residual. The displacement data can extract action speed, acceleration, displacement curve fitting degree, and displacement fluctuation variance. The flow change data can collect flow fluctuation coefficient and mutation jump point.

[0043] After the extracted features are standardized, a multi-dimensional feature vector matrix is constructed, each row of which corresponds to a time window and contains multi-dimensional pressure features, multi-dimensional vibration features, multi-dimensional temperature features, multi-dimensional displacement features, and multi-dimensional flow features. The row index corresponds to the time window and reflects the dynamic change of the features over time; the column index corresponds to the feature dimension, and each column represents the value sequence of a certain feature in all windows. The matrix can be directly used as the input of the AI model for subsequent state recognition and trend prediction, and the time continuity and multi-dimensionality of the matrix can completely retain the dynamic law of the performance degradation of the blowout preventer.

[0044] In the state recognition module, a state recognition model is constructed according to the multi-dimensional feature vector matrix to identify the current performance state of the blowout preventer and generate a probability distribution of the performance state. The state recognition model can be constructed in the following manner: a CNN-LSTM hybrid neural network architecture is constructed, in which the CNN is responsible for extracting the spatial patterns of the multi-dimensional features, the LSTM is responsible for capturing the time sequence dependency, and the two are fused to form a comprehensive feature representation; the blowout preventer state is divided into four levels of normal, slight abnormality, moderate abnormality, and severe abnormality according to industry standards and historical failure cases; a weighted cross-entropy loss function is used for dynamic sample weighting; historical operation data of the blowout preventer are collected to train and fine-tune the model, and a classification decision vector corresponding to the state is obtained.

[0045] In specific embodiments, the multi-dimensional feature vector matrix is subjected to spatial feature extraction by the CNN module. The input of the continuous 3 time window features is subjected to local correlation learning of the multi-dimensional features including pressure and vibration, such as capturing the synchronous pattern of "pressure mutation and vibration enhancement", and 16-dimensional feature maps with a size of [3, 14] are output; the second layer of convolution kernels further compresses the spatial dimension and strengthens the cross-feature interaction, such as integrating the correlation between the temperature change rate and the flow fluctuation, and 32 feature maps with a size of [2, 13] are output; finally, the spatial information is compressed into a 32-dimensional vector through global average pooling to form a spatial pattern feature, which contains the synergistic relationship of each dimension feature.

[0046] The LSTM module models the time sequence dependency of the spatial features. The 32-dimensional spatial features are input into a 2-layer bidirectional LSTM in the order of time windows: the first layer learns the short-term time sequence law, such as the lagging change of the pressure variance after the vibration entropy value increases in adjacent windows, and outputs 64-dimensional short-term time sequence features; the second layer captures the long-term dependency, such as the downward trend of the displacement fitting degree after the flow fluctuation coefficient is abnormal for 3 windows, and outputs 32-dimensional long-term time sequence features, which completely retain the dynamic evolution law of the features over time.

[0047] Subsequently, the spatial features and the time sequence features are spliced into a 64-dimensional comprehensive feature vector, and a full connection layer is used to complete classification mapping. The full connection layer uses a ReLU activation function to non-linearly map the 64-dimensional features to 4 neurons, which correspond to four states respectively. The output value of each neuron is the original score value. This process learns the association weight between the comprehensive features and each state through a weight matrix, and finally outputs four original score values, which reflect the matching degree between the input features and the corresponding state.

[0048] The probability distribution of the performance state is generated by converting the four original score values of the normal state, slight anomaly, moderate anomaly, and severe anomaly obtained by the model into a probability distribution through a Softmax function. The probability distribution can be expressed as: where P(y=k|z) is the probability of the device being in state k when the input feature z is input, z is the feature vector, w k is the classification decision vector corresponding to state k, is the matching degree between the feature and the state decision hyperplane,

[0049] The probability distribution is temperature scaled and corrected by introducing a temperature parameter T to adjust the probability distribution. The value of T is optimized through the validation set, for example, when T=0.7, the Brier score of the predicted probability and the actual state is the lowest, making the probability more consistent with the true uncertainty.

[0050] The uncertainty quantification is achieved by using the Monte Carlo dropout method, which enables the dropout layer during inference to repeatedly predict the same feature matrix and calculate the mean and standard deviation of the probability of each state to obtain the final probability distribution of the performance state.

[0051] In the trend prediction module, a trend prediction model is constructed based on the performance state probability distribution and the historical state sequence to predict the future performance degradation trajectory and the remaining useful life of the preventer, and generate a prediction result, which can be obtained by:

[0052] A quantifiable health degree sequence is constructed. The health degree is calculated based on the probability distribution output by the state recognition model to obtain a health degree value between 0 and 1, and then the historical health degree sequence is formed in time window order. To adapt to the model input, the sequence needs to be normalized and divided into samples by the sliding window method. Each sample contains the previous health degree value as input and the corresponding health degree value as output label to capture the short-term degradation trend.

[0053] Based on the health degree sequence, a trend prediction model is constructed using LSTM and attention mechanism architecture to generate the future health degradation trajectory. The model receives the historical health degree sequence, learns the long-term dependence through the stacked LSTM structure, assigns weights to the key time points through the attention layer, records the nodes that have a significant impact on future degradation, and finally outputs the future health curve through the fully connected layer. At the same time, in order to adapt to the linear degradation of the sealing wear and the sudden degradation of the hydraulic jam, a fault type embedding vector based on the state recognition result is added to the model to improve the prediction accuracy of non-linear trajectories.

[0054] Based on trajectory prediction, a deep residual network is introduced to learn the mapping relationship between features and remaining useful life. The input of ResNet integrates three types of key information, including the past window health degree mean, variance, and decline rate history health degree features, the current state features of the probability distribution of state recognition, and the trajectory prediction features of the slope and inflection point position of the future health curve, which together constitute the input vector. The network is composed of residual blocks, which alleviate the gradient vanishing problem through jump connection, and finally directly outputs the RUL estimate value through the fully connected layer. Physical constraints are added to the output layer to ensure that RUL≥0 and not exceed the remaining time of the design life of the preventer, avoiding unreasonable prediction values.

[0055] The model training uses the preventer full life cycle dataset, which covers the complete records of multiple devices from installation to failure. The true RUL label at each time point is the time interval from that time to the time when the health degree drops below the maintenance threshold. The training process uses a phased strategy, first freezes ResNet, trains the trajectory prediction model with 80% data, and then jointly trains the two models. Time series cross-validation is used to avoid the same device data appearing in the training set and test set. The loss function uses mean squared error with a penalty term, which increases the weight of samples with RUL prediction error exceeding 24 hours to strengthen long-term prediction accuracy. The optimizer uses AdamW and decays the learning rate by period until the RUL prediction error of the validation set is below the preset threshold. The final output prediction results include health degradation trajectory and RUL estimate value, with key influencing factors labeled to provide quantitative basis for maintenance planning.

[0056] In the decision warning module, a multi-level warning system is constructed based on the prediction results to generate maintenance decision schemes. The multi-level warning system is constructed as follows: based on RUL and health degree thresholds, three levels of warning are divided, when RUL is greater than m days and health degree is greater than x%, it is level one warning, maintain monitoring frequency and generate state report daily; when RUL is greater than n days and less than m days, and health degree is greater than y% and less than x%, it is level two warning, spare parts should be coordinated and downtime should be arranged while developing repair process; when RUL is less than n days and health degree is less than y%, the standby preventer should be started immediately, drilling operation should be stopped, and maintenance team should be dispatched synchronously.

[0057] In the learning optimization module, the AI model performance is continuously optimized according to the decision scheme and the actual maintenance result feedback. When the prediction deviation of a certain type of fault, such as the aging of the annular preventer rubber core, continuously exceeds the threshold value, the feature importance evaluation is first started: the SHAP value analysis model is used to analyze the decision basis of the fault, and if the absolute value of the SHAP value of the key feature is less than 0.1, it is determined that the feature is missing; at the same time, the PermutationImportance method is combined to randomly disturb the feature values and calculate the prediction accuracy drop, and the top 30% features such as the rubber core displacement fluctuation variance and the hydraulic oil viscosity change that contribute to the prediction of the fault are selected and supplemented to the original feature matrix. For example, for the problem of large prediction deviation of the rubber core aging, after adding the rubber core compression amount attenuation rate feature, the prediction accuracy of this type of fault can be improved.

[0058] For the feedback data accumulated daily, an incremental learning strategy is adopted to avoid full retraining, a frozen fine-tuning mechanism is constructed based on the pre-trained model parameters, the bottom convolutional layer and the LSTM unit for extracting general features in the CNN-LSTM model are frozen, and only the upper fully connected layer and the output layer are fine-tuned; at the same time, an elastic weight consolidation algorithm is introduced to consolidate the key parameters related to historical fault identification in the model, so as to prevent forgetting caused by learning of new data. When the data distribution deviates, the domain adaptation algorithm is used to reduce the distribution difference: first, the Wasserstein distance between the source domain and the target domain is calculated, and if the distance value exceeds the threshold value, the field adversarial neural network training is started, the field discriminator is added to the original CNN-LSTM model, and the feature extractor is forced to output general features that are not sensitive to the current field through adversarial learning; at the same time, the maximum mean difference constraint is performed on the key parameters, so that the feature distribution distance between the source domain and the target domain is reduced.

[0059] If the maintenance cost of the secondary warning is too high, the decision rule is optimized through reinforcement learning, an agent environment interaction framework is constructed, the current health degree, the remaining life prediction value and the spare parts inventory are taken as state inputs, and the actions including "maintaining monitoring", "pre-maintenance" and "delayed maintenance" are taken as optional actions, and the weighted sum of "maintenance cost + downtime loss - fault risk cost" is taken as the reward function. The deep deterministic policy gradient algorithm is used to train the agent, and the decision rule is iteratively optimized through simulated interaction with the environment.

[0060] Secondly, in the drawings of the disclosed embodiments, only the structures related to the disclosed embodiments are involved, and other structures can be referred to the general design, and the same embodiments and different embodiments of the present application can be combined with each other without conflict;

[0061] Finally: the above only for the preferred embodiments of the present application, and not for limiting the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application, should be included in the scope of protection of the present application.

Claims

1. An AI-based blowout preventer product performance monitoring system, characterized in that, The method comprises the following steps: a data acquisition module is used to acquire blowout preventer operation data and output standardized multi-dimensional time sequence data flow; a feature extraction module is used to extract features from the standardized multi-dimensional time sequence data flow and output a multi-dimensional feature vector matrix; the multi-dimensional feature vector matrix is generated by standardizing the extracted features and constructing a multi-dimensional feature vector matrix, each row of which corresponds to a time window and contains multi-dimensional pressure features, multi-dimensional vibration features, multi-dimensional temperature features, multi-dimensional displacement features and multi-dimensional flow features; a state recognition module is used to construct a state recognition model according to the multi-dimensional feature vector matrix, recognize the current performance state of the blowout preventer, and generate a probability distribution of the performance state; the state recognition model is constructed by constructing a CNN-LSTM hybrid neural network architecture, wherein the CNN is responsible for extracting the spatial mode of the multi-dimensional features, the LSTM is responsible for capturing the time sequence dependence, and the two are fused to form a comprehensive feature representation; the blowout preventer state is divided into four levels of normal, slight abnormality, moderate abnormality and severe abnormality according to industry standards and historical failure cases; a weighted cross-entropy loss function is used for dynamic sample weighting; historical operation data of the blowout preventer is collected to train and fine-tune the model, and a classification decision vector corresponding to the state is obtained; a trend prediction module is used to construct a trend prediction model according to the performance state probability distribution and the historical state sequence, predict the future performance degradation trajectory and the remaining useful life of the blowout preventer, and generate a prediction result; the performance state probability distribution is generated by converting the four original score values of the normal state, slight abnormality, moderate abnormality and severe abnormality obtained by the model into a probability distribution through a Softmax function, which can be expressed as: ; where P(y = k | z) is the probability of being in state k at the device given the input features z, z is the feature vector, w k is the classification decision vector for state k, is the degree of match of the features to the state decision hyperplane; the probability distribution is temperature-scaled and corrected and the uncertainty is quantified to obtain the final performance state probability distribution; the future performance degradation trajectory and the remaining useful life of the blowout preventer are predicted by constructing a trend prediction model, inputting a historical health degree sequence, and outputting a future health degree curve; on the basis of trajectory prediction, a deep residual network is used to learn the mapping relationship between features and RUL; the real RUL at each time point is labeled using the blowout preventer life cycle data, and the model is optimized in terms of mean square error; a future health degree degradation trajectory and an RUL estimate value are output; a decision warning module is used to construct a multi-level warning system according to the prediction result and generate a maintenance decision scheme; the multi-level warning system is constructed by dividing the RUL and health degree threshold into three levels of warning, when the RUL is greater than m days and the health degree is greater than x%, it is a first-level warning, the monitoring frequency is maintained, and a state report is generated daily; when the RUL is greater than n days and less than m days and the health degree is greater than y% and less than x%, it is a second-level warning, spare parts should be coordinated, downtime should be arranged, and a maintenance process should be developed; when the RUL is less than n days and the health degree is less than y%, a standby blowout preventer should be started immediately, drilling operations should be stopped, and a maintenance team should be dispatched synchronously; a learning optimization module is used to continuously optimize the AI model performance according to the decision scheme and actual maintenance result feedback. The feature extraction includes: hydraulic system pressure data extraction dynamic pressure mean and variance, pressure mutation index, FFT main frequency energy proportion, harmonic distortion rate; vibration data extraction resonance frequency band energy, fault characteristic frequency, wavelet energy entropy; temperature change data extraction temperature change rate, cumulative heat effect, moving average residual; displacement data extraction action speed, acceleration, displacement curve fitting degree, displacement fluctuation variance; flow change data collection flow fluctuation coefficient, mutation jump point.

2. The AI-based blowout preventer product performance monitoring system of claim 1, wherein: The blowout preventer operation data includes: pressure sensor collects hydraulic system pressure data; vibration sensor monitors mechanical component vibration data; temperature sensor monitors key position temperature change data; displacement sensor detects ram and annular blowout preventer action displacement data; flow sensor monitors hydraulic oil flow change data.

3. The AI-based blowout preventer product performance monitoring system of claim 1, wherein: The way to continuously optimize the AI model performance is: if the prediction deviation of a certain type of fault is large, analyze whether the key features are missed, and use feature importance evaluation to screen effective features; use feedback data for incremental learning to avoid full volume retraining; for distribution offset data, use domain adaptation algorithm to reduce distribution difference; if the maintenance cost of secondary early warning is too high, adjust the decision rule and optimize the decision strategy through reinforcement learning.

4. An AI-based blowout preventer product performance monitoring method applied to the AI-based blowout preventer product performance monitoring system of any one of claims 1-3, characterized in that, It includes: S1, collecting blowout preventer operation data, outputting standardized multi-dimensional time series data stream; S2, performing feature extraction on the standardized multi-dimensional time series data stream to output a multi-dimensional feature vector matrix; S3, constructing a state recognition model according to the multi-dimensional feature vector matrix to recognize the current performance state of the blowout preventer and generate a probability distribution of the performance state; S4, constructing a trend prediction model according to the performance state probability distribution and historical state sequence to predict the future performance degradation trajectory and remaining useful life of the blowout preventer and generate a prediction result; S5, constructing a multi-level early warning system according to the prediction result to generate a maintenance decision scheme; S6, continuously optimizing the AI model performance according to the decision scheme and actual maintenance result feedback.

Images