Blowout preventer product performance monitoring system and method based on AI

Abstract

The invention relates to the technical field of blowout preventer monitoring, in particular to an AI-based blowout preventer product performance monitoring system and method, and the system comprises a data collection module which is used for collecting the operation data of a blowout preventer and outputting a standardized multi-dimensional time sequence data stream; the feature extraction module is used for performing feature extraction on the standardized multi-dimensional time sequence data stream and outputting a multi-dimensional feature vector matrix; the state recognition module 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 probability distribution of the performance state; the trend prediction module is used for constructing a trend prediction model according to the performance state probability distribution and the historical state sequence, predicting the future performance degradation track and the remaining service life of the blowout preventer and generating a prediction result; the decision early warning module is used for constructing a multi-level early warning system according to the prediction result and generating a maintenance decision scheme; and the learning optimization module is used for continuously optimizing the AI model performance according to the decision scheme and the actual maintenance result feedback.

Description

Technical Field

[0001] This invention relates to the field of blowout preventer monitoring technology, specifically an AI-based blowout preventer product performance monitoring system and method. Background Technology

[0002] Traditional blowout preventer (BOP) performance monitoring solutions rely on manual inspections and single-parameter monitoring during the data acquisition phase. Typically, only simple equipment is used to record key pressure values, resulting in low sampling frequency and a lack of multi-dimensional data such as vibration and temperature. Data processing depends on manual entry into spreadsheets, which is prone to errors and omissions, and cannot generate standardized time-series data. This method struggles to capture instantaneous fault signals, such as pressure fluctuations caused by sudden wear of seals, which are often only discovered after the fault has occurred. In the status identification and early warning stages, traditional solutions rely on fixed thresholds for judgment. Maintenance personnel set safe ranges for pressure and temperature parameters according to industry standards; exceeding these ranges triggers an alarm. However, blowout preventer performance degradation is a gradual process, and fixed thresholds cannot identify early, minor anomalies, such as vibration spectrum shifts in the early stages of gate wear. Furthermore, the lack of a tiered alarm mechanism leads to slow responses from maintenance personnel to critical warnings, increasing the risk of failure. In terms of trend prediction and maintenance decisions, traditional solutions rely on experience-based judgment. Engineers develop fixed-cycle preventative maintenance plans based on the equipment's age and historical failure records, such as mandatory maintenance every six months. This one-size-fits-all approach often leads to over-maintenance or under-maintenance, increasing costs and failing to accurately address the performance differences of individual equipment. Traditional solutions lack dynamic optimization mechanisms. Once the models and decision-making logic are determined, they remain unchanged for a long period, unable to adapt to factors such as blowout preventer aging and changes in operating conditions. For example, the performance implications of new equipment and equipment that has been used for 5 years are different under the same parameters, but traditional solutions cannot dynamically adjust the judgment criteria. At the same time, the deviation between maintenance results and predictions is not used to improve the system, making it difficult to improve monitoring accuracy, and remaining at the level of passive response rather than proactive prediction. Summary of the Invention

[0003] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: an AI-based blowout preventer product performance monitoring system, characterized in that it includes: The data acquisition module is used to collect the operation data of the blowout preventer and output a standardized multi-dimensional time-series data stream; The feature extraction module is used to extract features from a standardized multidimensional time-series data stream and output a multidimensional feature vector matrix. The status recognition module is used to construct a status recognition model based on a multi-dimensional feature vector matrix to identify the current performance status of the blowout preventer and generate a probability distribution of the performance status. The trend prediction module is used to build a trend prediction model based on the performance status probability distribution and historical status sequence, predict the future performance degradation trajectory and remaining service life of the blowout preventer, and generate prediction results. The decision-making and early warning module is used to build a multi-level early warning system based on the prediction results and generate maintenance decision-making schemes. The learning optimization module is used to continuously optimize the performance of the AI ​​model based on the decision-making process and feedback from actual maintenance results.

[0004] The blowout preventer operating data includes: pressure sensor collecting hydraulic system pressure data; vibration sensor monitoring mechanical component vibration data; temperature sensor monitoring temperature change data of key parts; displacement sensor detecting the displacement data of the gate and annular blowout preventer; and flow sensor monitoring hydraulic oil flow change data.

[0005] The feature extraction includes: extracting dynamic pressure mean and variance, pressure mutation index, FFT main frequency energy ratio, and harmonic distortion rate from hydraulic system pressure data; extracting resonant frequency band energy, fault characteristic frequency, and wavelet energy entropy from vibration data; extracting temperature change rate, cumulative thermal effect, and moving average residual from temperature change data; extracting action velocity, acceleration, displacement curve fitting degree, and displacement fluctuation variance from displacement data; and collecting flow fluctuation coefficient and mutation jump point from flow change data.

[0006] The specific process of generating the multidimensional feature vector matrix is ​​as follows: after standardizing the extracted features, a multidimensional feature vector matrix is ​​constructed, with each row corresponding to a time window, including multidimensional pressure features, multidimensional vibration features, multidimensional temperature features, multidimensional displacement features, and multidimensional flow features.

[0007] The method for constructing the state recognition model is as follows: a CNN-LSTM hybrid neural network architecture is constructed, in which CNN is responsible for extracting spatial patterns of multi-dimensional features, LSTM is responsible for capturing temporal dependencies, and the two are fused to form a comprehensive feature representation; the blowout preventer state is divided into four levels: normal, slightly abnormal, moderately abnormal, and severely abnormal, based on industry standards and historical failure cases; a weighted cross-entropy loss function is used for dynamic sample weighting; historical operating data of the blowout preventer is collected to train and fine-tune the model, and the classification decision vector for the corresponding state is obtained.

[0008] The probability distribution of the generated performance states is obtained by converting the four original scores (normal state, slight anomaly, moderate anomaly, and severe anomaly) from the model into a probability distribution using the Softmax function, which can be expressed as: Where P(y=k|z) is the probability that the device is in state k when input feature z, z is the feature vector, and w k Let k be the classification decision vector corresponding to state k. The degree of matching between features and the state-decision hyperplane; Temperature scaling correction and uncertainty quantification are applied to the probability distribution to obtain the probability distribution of the final performance state.

[0009] The method for predicting the future performance degradation trajectory and remaining service life of the blowout preventer is as follows: a trend prediction model is constructed, a historical health sequence is input, and a future health curve is output; based on the trajectory prediction, a deep residual network is used to learn the mapping relationship between features and RUL; the true RUL at each time point is labeled using the full life cycle data of the blowout preventer, and the model is optimized with mean square error; the future health degradation trajectory and RUL estimate are output.

[0010] The method for constructing a multi-level early warning system is as follows: Based on RUL and health thresholds, three levels of early warning are established. When RUL is greater than m days and health is greater than x%, it is a Level 1 early warning, maintaining monitoring frequency and generating a daily status report. When RUL is greater than n days but less than m days and health is greater than y% but less than x%, it is a Level 2 early warning, requiring coordination of spare parts, scheduling of downtime, and the development of maintenance procedures. When RUL is less than n days and health is less than y%, the backup blowout preventer should be activated immediately, drilling operations stopped, and the maintenance team dispatched simultaneously.

[0011] The method for continuously optimizing the performance of the AI ​​model is as follows: if the prediction deviation of a certain type of fault is large, analyze whether key features are missed, and use feature importance assessment to screen effective features; use feedback data for incremental learning to avoid full retraining; for data with distribution deviation, use domain adaptation algorithm to reduce distribution difference; if the maintenance cost of secondary early warning is too high, adjust the decision rules and optimize the decision strategy through reinforcement learning.

[0012] A method for monitoring the performance of blowout preventer products based on AI, characterized by comprising: S1. Collect blowout preventer operation data and output a standardized multi-dimensional time-series data stream; S2. Extract features from the standardized multi-dimensional time-series data stream and output a multi-dimensional feature vector matrix; S3. Based on the multidimensional feature vector matrix, construct a state recognition model to identify the current performance state of the blowout preventer and generate a probability distribution of the performance state. S4. Based on the performance status probability distribution and historical status sequence, construct a trend prediction model to predict the future performance degradation trajectory and remaining service life of the blowout preventer, and generate prediction results. S5. Construct a multi-level early warning system based on the prediction results and generate maintenance decision-making plans; S6. Continuously optimize AI model performance based on decision-making plans and actual maintenance results.

[0013] Combining all the above technical solutions, the positive effects of this invention are as follows: 1. This invention breaks through the limitations of information silos in traditional monitoring through multi-dimensional data acquisition and standardized processing. By integrating data from multiple sensors such as pressure, vibration, and temperature, and processing it through noise reduction and normalization to form a time-series data stream, it achieves holographic perception of the blowout preventer's operating status. Compared with the single-point sampling of traditional manual inspections, this method can capture transient anomalies, improve data integrity, lay a high-quality foundation for subsequent AI analysis, and solve the problem of missed judgments caused by incomplete data in traditional solutions from the source.

[0014] 2. In terms of state recognition and trend prediction, this invention achieves a leap from passive alarm to proactive prediction. By outputting a performance state probability distribution through a state recognition model, it accurately distinguishes between slight degradation and normal conditions; combined with historical sequence prediction of degradation trajectory and remaining lifetime, the prediction error is reduced compared to traditional empirical estimation. This early warning capability can avoid the risk of blowouts caused by sudden failures, while providing precise timing for planned shutdowns, significantly improving drilling operation safety.

[0015] 3. This invention significantly optimizes operation and maintenance efficiency through a multi-level early warning and intelligent decision-making mechanism. Based on prediction results, tiered responses are triggered, forming a closed-loop management system from enhanced monitoring to emergency shutdown, reducing the false alarm storms associated with traditional fixed-threshold alarms. Maintenance plans generated based on operating conditions can balance spare parts inventory with drilling progress, avoiding resource waste caused by over-maintenance. In practical applications, this can shorten unplanned downtime, reduce annual maintenance costs, and balance safety and economy.

[0016] 4. This invention ensures the long-term effectiveness of the system through a continuous learning optimization mechanism. By dynamically updating the model based on maintenance results, it adapts to complex scenarios such as blowout preventer aging and changing operating conditions, addressing the pain point of poor adaptability caused by fixed parameters in traditional solutions. With data accumulation, the accuracy of status recognition can be gradually improved, and the accuracy of remaining life prediction can be continuously optimized, forming a positive cycle of data, model, and decision-making, providing intelligent support for the full lifecycle management of blowout preventers. Attached Figure Description

[0017] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0018] Figure 1 This is a system framework diagram of the system of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] See Figure 1 As shown, this invention proposes an AI-based performance monitoring system for blowout preventers, comprising: a data acquisition module, a feature extraction module, a status recognition module, a trend prediction module, a decision warning module, and a learning optimization module.

[0021] In a more specific application of this invention, the data acquisition module uses pressure sensors to collect hydraulic system pressure data. High-precision pressure sensors are installed at key nodes in the blowout preventer's hydraulic control circuit, including the gate cylinder inlet, the annular blowout preventer control pipeline, and the accumulator outlet. The sensors are connected to the hydraulic pipeline via threaded connections to sense dynamic pressure changes in the hydraulic oil in real time, with a sampling frequency of 10Hz. The collected raw data is converted into a 4-20mA current signal by a signal conditioning module, transmitted to an edge computing gateway, where digital filtering removes spike noise generated by hydraulic shocks. Finally, the data is stored in a time-series database in time-pressure format, forming a continuous monitoring sequence of hydraulic system pressure.

[0022] Vibration sensors monitoring the vibration data of mechanical components utilize piezoelectric accelerometers, which are magnetically or bolted to the surfaces of key mechanical components such as the blowout preventer housing, gate drive shaft, and annular rubber core cover. The sensors acquire vibration acceleration signals along the X, Y, and Z axes in three dimensions, with a sampling frequency up to 1 kHz to capture high-frequency impact signals. The raw vibration signals are converted into voltage signals by a charge amplifier and then uploaded in real-time via industrial Ethernet. Wavelet threshold denoising is then applied to eliminate environmental vibration interference from the drilling platform while retaining characteristic frequency components of mechanical faults, providing high signal-to-noise ratio data for subsequent feature extraction.

[0023] Temperature sensors monitor temperature changes in key components using a combination of armored thermocouples and infrared temperature sensors. The thermocouples are embedded in contact measuring points on the gate sealing surface and the inner wall of the hydraulic oil lines, transmitting temperature signals via compensating wires. The infrared sensor is installed 20-30 cm away from the annular rubber core, providing non-contact monitoring of the core's surface temperature. The sampling frequency is set to 1 Hz, and the data, after cold junction compensation and linearization, is time-aligned with pressure and vibration data. When the temperature change rate exceeds 0.5℃ / min, the sampling frequency is automatically increased to 10 Hz to focus on capturing abnormal temperature rise processes.

[0024] Displacement sensors detect 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 cylinder. The sensor moves synchronously with the piston rod via an internal magnetic ring, outputting the absolute displacement value of the gate's opening and closing stroke in real time. The annular blowout preventer uses a wire-type displacement sensor, with the wire end connected to the piston core. The compression of the piston core is calculated by the rotation angle of the spool. The sensor outputs a 4-20mA analog signal, which, after A / D conversion, is used to verify its validity in conjunction with the blowout preventer's action command. Abnormal jump values ​​caused by mechanical jamming are eliminated, ultimately forming a displacement-time curve that visually reflects the accuracy of the action.

[0025] A flow sensor monitors changes in hydraulic oil flow rate by installing turbine flow meters at the hydraulic pump outlet and the main inlet line of the blowout preventer. The rotating turbine blades cut magnetic lines of force to generate pulse signals, with the pulse frequency proportional to the flow rate. The sensor's output pulse signal is converted into an instantaneous flow rate value by a counter, with a sampling frequency of 5Hz. Simultaneously, the total flow rate per unit time is calculated. By comparing the flow rate difference between the inlet and return lines, the presence of internal leaks in the hydraulic system can be determined. Combined with pressure data correlation analysis, this improves the accuracy of fault location. In the feature extraction module, standardized multidimensional time-series data streams are processed to extract features, outputting a multidimensional feature vector matrix. This can be achieved through: extracting dynamic pressure mean and variance, pressure mutation index, FFT main frequency energy proportion, and harmonic distortion rate from hydraulic system pressure data; extracting resonant frequency band energy, fault characteristic frequency, and wavelet energy entropy from vibration data; extracting temperature change rate, cumulative thermal effect, and moving average residual from temperature change data; extracting action velocity, acceleration, displacement curve fitting degree, and displacement fluctuation variance from displacement data; and acquiring flow fluctuation coefficient and mutation jump points from flow change data.

[0026] After standardizing the extracted features, a multidimensional feature vector matrix is ​​constructed. Each row corresponds to a time window, containing multidimensional pressure, vibration, temperature, displacement, and flow characteristics. The row index corresponds to the time window, reflecting the dynamic changes of the features over time; the column index corresponds to the feature dimension, with each column representing the value sequence of a certain feature across all windows. This matrix can be directly used as input to an AI model for subsequent state recognition and trend prediction. The temporal continuity and multidimensionality of the matrix can fully preserve the dynamic laws governing the performance degradation of the blowout preventer.

[0027] In the status recognition module, a status recognition model is constructed based on the multi-dimensional feature vector matrix to identify the current performance status of the blowout preventer and generate a probability distribution of the performance status. This can be achieved by constructing a CNN-LSTM hybrid neural network architecture, where CNN is responsible for extracting the spatial patterns of multi-dimensional features, and LSTM is responsible for capturing temporal dependencies. The two are then fused to form a comprehensive feature representation. Combining industry standards and historical failure cases, the blowout preventer status is divided into four levels: normal, slightly abnormal, moderately abnormal, and severely abnormal. A weighted cross-entropy loss function is used for dynamic sample weighting. Historical operating data of the blowout preventer is collected to train and fine-tune the model, while simultaneously obtaining the classification decision vector for the corresponding status.

[0028] In a specific embodiment, the multidimensional feature vector matrix is ​​processed by a CNN module to extract spatial features. The input features from three consecutive time windows are used to perform local correlation learning on multidimensional features including pressure and vibration, such as capturing the synchronous pattern of "pressure change and vibration enhancement", and outputting 16 feature maps with dimensions [3,14]. The second convolutional kernel further compresses the spatial dimensions and strengthens cross-feature interaction, such as fusing the correlation between temperature change rate and flow fluctuation, and outputs 32 feature maps [2,13]. Finally, global average pooling is used to compress the spatial information into a 32-dimensional vector to form spatial pattern features, which contain the synergistic relationship of features in each dimension.

[0029] The LSTM module models the temporal dependencies of spatial features. 32-dimensional spatial features are input sequentially into a two-layer bidirectional LSTM: the first layer learns short-term temporal patterns through forward and backward propagation, such as the "hysteretic change in pressure variance after an increase in vibration entropy" in adjacent windows, outputting 64-dimensional short-term temporal features; the second layer captures long-term dependencies, such as the "decreasing trend of displacement fit after an abnormal flow fluctuation coefficient persists for three windows," outputting 32-dimensional long-term temporal features, fully preserving the dynamic evolution of features over time.

[0030] Subsequently, spatial and temporal features are concatenated into a 64-dimensional comprehensive feature vector, which is then processed by a fully connected layer for classification mapping. The fully connected layer employs the ReLU activation function, non-linearly mapping the 64-dimensional features to four neurons, each corresponding to one of the four states. The output value of each neuron is the original score. This process learns the association weights between the comprehensive features and each state through a weight matrix, ultimately outputting four original scores, the magnitude of which reflects the degree of matching between the input features and the corresponding states.

[0031] The probability distribution for generating performance states is achieved by converting the four raw scores (normal, slightly abnormal, moderately abnormal, and severely abnormal) obtained from the model into a probability distribution using the Softmax function, as expressed by: Where P(y=k|z) is the probability that the device is in state k when input feature z, z is the feature vector, and wk Let k be the classification decision vector corresponding to state k. The degree of matching between features and the state-decision hyperplane; Temperature scaling correction is applied to the probability distribution by introducing a temperature parameter T to adjust the probability distribution. The T value is optimized through the validation set. For example, when T=0.7, the Brier score of the predicted probability is the lowest, making the probability more consistent with the real uncertainty.

[0032] Uncertainty quantification employs the Monte Carlo dropout method, which enables a dropout layer during inference, repeatedly predicts the same feature matrix, calculates the mean and standard deviation of the probabilities of each state, and obtains the probability distribution of the final performance state.

[0033] In the trend prediction module, a trend prediction model is constructed based on the performance state probability distribution and historical state sequence to predict the future performance degradation trajectory and remaining service life of the blowout preventer, generating prediction results that can be accessed through: A quantifiable health sequence is constructed, with health values ​​calculated based on the probability distribution output by the state recognition model, resulting in health values ​​between 0 and 1. These values ​​are then concatenated according to time windows to form a historical health sequence. To adapt to the model input, the sequence needs to be normalized, and samples are divided using a sliding window method. Each sample contains the previous health value as input and the corresponding subsequent health value as the output label, thereby capturing short-term degradation trends.

[0034] Based on health status sequences, a trend prediction model is constructed using an LSTM and attention mechanism architecture to generate future health status degradation trajectories. The model receives historical health status sequences, learns long-term dependencies through a stacked LSTM structure, assigns weights to key time points using an attention layer, and records nodes that significantly impact future degradation. Finally, a fully connected layer outputs the future health status curve. To adapt to different fault types, such as linear degradation from seal wear and abrupt degradation from hydraulic jamming, a fault type embedding vector based on state recognition results is added to the model, improving the prediction accuracy for nonlinear trajectories.

[0035] Building upon trajectory prediction, a deep residual network is introduced to learn the mapping relationship between features and remaining service life. The ResNet input integrates three key types of information: historical health features such as the mean, variance, and rate of decline of past health conditions; current state features from the probability distribution of state recognition; and trajectory prediction features such as the slope and inflection point position of the future health curve, collectively forming the input vector. The network consists of residual blocks, and skip connections mitigate the vanishing gradient problem. Finally, a fully connected layer directly outputs the estimated RUL value, with physical constraints added to the output layer to ensure that RUL ≥ 0 and does not exceed the remaining time of the blowout preventer's design life, thus avoiding unreasonable predictions.

[0036] The model training utilizes a full lifecycle dataset for blowout preventers, encompassing complete records of multiple devices from installation to failure. The true Recovery Usage (RUL) at each time point is labeled as the interval between that time and the point at which the health status falls below the maintenance threshold. A phased training strategy is employed: first, ResNet is frozen, and the trajectory prediction model is trained using 80% of the data. Then, the two models are jointly trained, employing time-series cross-validation to prevent data from the same device from appearing simultaneously in both the training and test sets. The loss function uses mean squared error plus a penalty term, with increased weights for samples with RUL prediction errors exceeding 24 hours to enhance long-term prediction accuracy. The optimizer uses AdamW with a periodically decaying learning rate until the RUL prediction error on the validation set falls below a preset threshold. The final output includes the health degradation trajectory and estimated RUL values, along with annotations of key influencing factors, providing a quantitative basis for maintenance planning.

[0037] In the decision-making and early warning module, a multi-level early warning system is constructed based on the prediction results to generate maintenance decision-making plans. The multi-level early warning system can be constructed as follows: Based on RUL and health threshold, three levels of early warning are divided. When RUL is greater than m days and health is greater than x%, it is a level one early warning, and the monitoring frequency is maintained and a status report is generated daily. When RUL is greater than n days but less than m days and health is greater than y% but less than x%, it is a level two early warning, and spare parts should be coordinated, downtime should be arranged, and maintenance procedures should be formulated. When RUL is less than n days and health is less than y%, the backup blowout preventer should be activated immediately, drilling operations should be stopped, and the maintenance team should be dispatched simultaneously.

[0038] In the learning and optimization module, the performance of the AI ​​model is continuously optimized based on the decision-making scheme and feedback from actual maintenance results. This can be achieved by: when the prediction deviation of a certain type of fault, such as the aging of the rubber core of the annular blowout preventer, consistently exceeds a threshold, firstly, a feature importance assessment is initiated: the SHAP value is used to analyze the model's decision-making basis for this type of fault. If the absolute value of the SHAP value of a key feature is less than 0.1, it is determined to be a feature omission; simultaneously, the PermutationImportance method is used to randomly shuffle the feature values ​​and calculate the decrease in prediction accuracy, selecting the top 30% of features that contribute to the prediction of this fault, such as the variance of rubber core displacement fluctuation and hydraulic oil viscosity change, and adding them to the original feature matrix. For example, to address the problem of large prediction deviations in rubber core aging, adding the feature of rubber core compression attenuation rate can improve the prediction accuracy of this type of fault.

[0039] For the daily accumulated feedback data, an incremental learning strategy is adopted to avoid full retraining. A freeze-fine-tuning mechanism is constructed based on the pre-trained model parameters, freezing the bottom convolutional layers and LSTM units that extract general features from the CNN-LSTM model, and only fine-tuning the upper fully connected layers and output layers. At the same time, an elastic weight consolidation algorithm is introduced to apply penalty terms to key parameters in the model related to historical fault identification to prevent forgetting caused by learning new data. When a data distribution shift is detected, the distribution difference is reduced through a domain adaptation algorithm: first, the Wasserstein distance between the source domain and the target domain is calculated. If the distance value exceeds a threshold, the training of the domain adversarial neural network is started, and a domain discriminator is added to the original CNN-LSTM model. Through adversarial learning, the feature extractor is forced to output general features that are not sensitive to the current domain. At the same time, the maximum mean difference constraint is applied to the key parameters to reduce the feature distribution distance between the source domain and the target domain.

[0040] If the maintenance cost of a Level 2 early warning system is too high, reinforcement learning is used to optimize the decision-making rules and construct an intelligent agent-environment interaction framework. The current health status, predicted remaining lifespan, and spare parts inventory are used as state inputs. "Maintain monitoring," "early maintenance," and "delayed maintenance" are set as optional actions. The reward function is a weighted sum of "maintenance cost + downtime loss - failure risk cost." A deep deterministic policy gradient algorithm is used to train the agent, and the decision-making rules are iteratively optimized through simulated interaction with the environment.

[0041] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI-based performance monitoring system for blowout preventers, characterized in that, include: The data acquisition module is used to collect the operation data of the blowout preventer and output a standardized multi-dimensional time-series data stream; The feature extraction module is used to extract features from a standardized multidimensional time-series data stream and output a multidimensional feature vector matrix. The status recognition module is used to construct a status recognition model based on a multi-dimensional feature vector matrix to identify the current performance status of the blowout preventer and generate a probability distribution of the performance status. The trend prediction module is used to build a trend prediction model based on the performance status probability distribution and historical status sequence, predict the future performance degradation trajectory and remaining service life of the blowout preventer, and generate prediction results. The decision-making and early warning module is used to build a multi-level early warning system based on the prediction results and generate maintenance decision-making schemes. The learning optimization module is used to continuously optimize the performance of the AI ​​model based on the decision-making plan and feedback from actual maintenance results; The feature extraction includes: extracting dynamic pressure mean and variance, pressure mutation index, FFT main frequency energy ratio, and harmonic distortion rate from hydraulic system pressure data; extracting resonant frequency band energy, fault characteristic frequency, and wavelet energy entropy from vibration data; extracting temperature change rate, cumulative thermal effect, and moving average residual from temperature change data; extracting action velocity, acceleration, displacement curve fitting degree, and displacement fluctuation variance from displacement data; and collecting flow fluctuation coefficient and mutation jump point from flow change data.

2. The AI-based blowout preventer product performance monitoring system as described in claim 1, characterized in that: The blowout preventer operating data includes: pressure sensor collecting hydraulic system pressure data; vibration sensor monitoring mechanical component vibration data; temperature sensor monitoring temperature change data of key parts; displacement sensor detecting the displacement data of the gate and annular blowout preventer; and flow sensor monitoring hydraulic oil flow change data.

3. The AI-based blowout preventer product performance monitoring system as described in claim 1, characterized in that: The specific method for generating the multidimensional feature vector matrix is ​​as follows: after standardizing the extracted features, a multidimensional feature vector matrix is ​​constructed, with each row corresponding to a time window, including multidimensional pressure features, multidimensional vibration features, multidimensional temperature features, multidimensional displacement features, and multidimensional flow features.

4. The AI-based blowout preventer product performance monitoring system as described in claim 1, characterized in that: The method for constructing the state recognition model is as follows: a CNN-LSTM hybrid neural network architecture is constructed, in which CNN is responsible for extracting spatial patterns of multi-dimensional features, LSTM is responsible for capturing temporal dependencies, and the two are fused to form a comprehensive feature representation; the blowout preventer state is divided into four levels: normal, slightly abnormal, moderately abnormal, and severely abnormal, based on industry standards and historical failure cases; a weighted cross-entropy loss function is used for dynamic sample weighting; historical operating data of the blowout preventer is collected to train and fine-tune the model, and the classification decision vector for the corresponding state is obtained.

5. The AI-based blowout preventer product performance monitoring system as described in claim 1, characterized in that: The probability distribution of the generated performance states is obtained by converting the four original scores (normal state, slight anomaly, moderate anomaly, and severe anomaly) from the model into a probability distribution using the Softmax function, which can be expressed as: Where P(y=k|z) is the probability that the device is in state k when input feature z, z is the feature vector, and w k Let k be the classification decision vector corresponding to state k. The degree of matching between features and the state-decision hyperplane; Temperature scaling correction and uncertainty quantification are applied to the probability distribution to obtain the probability distribution of the final performance state.

6. The AI-based blowout preventer product performance monitoring system as described in claim 1, characterized in that: The method for predicting the future performance degradation trajectory and remaining service life of the blowout preventer is as follows: a trend prediction model is constructed, a historical health sequence is input, and a future health curve is output; based on the trajectory prediction, a deep residual network is used to learn the mapping relationship between features and RUL; the true RUL at each time point is labeled using the full life cycle data of the blowout preventer, and the model is optimized with mean square error; the future health degradation trajectory and RUL estimate are output.

7. The AI-based blowout preventer product performance monitoring system as described in claim 1, characterized in that: The method for constructing a multi-level early warning system is as follows: Based on RUL and health thresholds, three levels of early warning are established. When RUL is greater than m days and health is greater than x%, it is a Level 1 early warning, maintaining monitoring frequency and generating a daily status report. When RUL is greater than n days but less than m days and health is greater than y% but less than x%, it is a Level 2 early warning, requiring coordination of spare parts, scheduling of downtime, and the development of maintenance procedures. When RUL is less than n days and health is less than y%, the backup blowout preventer should be activated immediately, drilling operations stopped, and the maintenance team dispatched simultaneously.

8. The AI-based blowout preventer product performance monitoring system as described in claim 1, characterized in that: The method for continuously optimizing the performance of the AI ​​model is as follows: if the prediction deviation of a certain type of fault is large, analyze whether key features are missed, and use feature importance assessment to screen effective features; use feedback data for incremental learning to avoid full retraining; for data with distribution deviation, use domain adaptation algorithm to reduce distribution difference; if the maintenance cost of secondary early warning is too high, adjust the decision rules and optimize the decision strategy through reinforcement learning.

9. A method for monitoring the performance of blowout preventer products based on AI, characterized in that, include: S1. Collect blowout preventer operation data and output a standardized multi-dimensional time-series data stream; S2. Extract features from the standardized multi-dimensional time-series data stream and output a multi-dimensional feature vector matrix; S3. Based on the multidimensional feature vector matrix, construct a state recognition model to identify the current performance state of the blowout preventer and generate a probability distribution of the performance state. S4. Based on the performance status probability distribution and historical status sequence, construct a trend prediction model to predict the future performance degradation trajectory and remaining service life of the blowout preventer, and generate prediction results. S5. Construct a multi-level early warning system based on the prediction results and generate maintenance decision-making plans; S6. Continuously optimize AI model performance based on decision-making plans and actual maintenance results.

Images