KR102618023B1 - Failure prediction diagnosis system and method through pattern analysis according to failure type - Google Patents

Abstract

The failure prediction diagnosis system and method through pattern analysis according to the failure type according to the embodiment provides a failure prediction diagnosis system through pattern analysis according to the failure type of oil and gas plant core equipment. In the embodiment, data is analyzed using machine learning and deep learning for failure prediction and diagnosis, and the failure rate is estimated through an artificial neural network model to diagnose device status or failure to predict the time of failure. We propose a foresight model. The diagnosis model can diagnose conditions or faults using a supervised learning SVM (Support Vector Machine)-based diagnosis model and an unsupervised learning One-Class SVM-based diagnosis model.

Description

Failure prediction diagnosis system and method through pattern analysis according to failure type {FAILURE PREDICTION DIAGNOSIS SYSTEM AND METHOD THROUGH PATTERN ANALYSIS ACCORDING TO FAILURE TYPE}

This disclosure relates to a failure prediction diagnosis system and method. Specifically, it relates to a failure prediction diagnosis system and method through pattern analysis according to failure type of oil and gas plant core equipment.

Unless otherwise indicated herein, the material described in this section is not prior art to the claims of this application, and is not admitted to be prior art by inclusion in this section.

Facilities and equipment composed of various parts may experience changes in temperature or pressure or increase in vibration and noise when problems begin to occur during operation.

do. and defects or damage due to failure if appropriate measures are not taken.

This can lead to shutdowns or accidents, resulting in social and economic losses. As a result, preventive maintenance through regular inspection or replacement of parts is carried out at regular intervals before a breakdown occurs to prevent breakdowns in advance, but the problem is that unnecessary replacement of parts causes cost loss and does not prevent sudden breakdowns. There is. To solve these problems, research on RCM (Reliability Centered Maintenance) and RBM (Risk-Based Maintenance), which prevent failures based on reliability, has been conducted, but field application is limited due to difficulties in coping with overall situations that occur during operation. .

Recently, with the development of ICT (Information and Communication Technology), PHM (Prognostics and Health Management), which collects data in real time by attaching various sensors to facilities and equipment, diagnoses failures, and predicts failure times through analysis techniques, is being studied. . PHM continuously monitors defects or performance degradation of facilities and equipment in operation, diagnoses abnormal signs, predicts the level or timing of failure, prevents losses resulting from breakdown through predictive maintenance, and ensures safety and availability of maintenance. This is a technology that improves operational efficiency. And for the timing of predictive maintenance, PHM approaches data based on probability statistics and measures RUL (Remaining Useful Life), which is the effective time before failure occurs. PHM can improve maintainability, economic efficiency, reliability, and availability over preventive maintenance, so research is being conducted domestically, but it is still insufficient.

1. Korean Patent Registration No. 10-2419782 (2022.07.07) 2. Korean Patent Registration No. 10-1987365 (2019.06.03)

The failure prediction diagnosis system and method through pattern analysis according to the failure type according to the embodiment provides a failure prediction diagnosis system through pattern analysis according to the failure type of oil and gas plant core equipment.

In the embodiment, data is analyzed using machine learning and deep learning for failure prediction and diagnosis, and the failure rate is estimated through an artificial neural network model to diagnose device status or failure to predict the time of failure. We propose a foresight model. The diagnosis model can diagnose conditions or faults using a supervised learning SVM (Support Vector Machine)-based diagnosis model and an unsupervised learning One-Class SVM-based diagnosis model.

The health prediction model provided in the embodiment includes a Markov Chain Monte Carlo (MCMC)-based prediction model that analyzes Time To Event (TTE) data in a non-parametric method and a Long Short-Term Memory (LSTM) model that analyzes data in a parametric method. Using a predictive model, the failure rate over time can be estimated and the time of failure can be predicted by measuring the remaining useful life (RUL) before failure occurs.

In addition, the diagnostic model and health prediction model provided in the embodiment allow the selection and application of an appropriate model according to the state of the data, and the timing of predictive maintenance can be determined by diagnosing the state or estimating the failure rate.

In an embodiment, a deep learning neural network including at least one of a Deep Neural Network (DNN), a Convolutional Neural Network (CNN), a Recurrent Neural Network (RNN), a Bidirectional Recurrent Deep Neural Network (BRDNN), and a Bayesian network. A failure prediction model is implemented by learning with a training data set for failure diagnosis.

The oil and gas plant equipment failure prediction system according to the embodiment includes a sensor installed inside the oil and gas plant to sense plant equipment operation information; A failure prediction server that collects and analyzes plant equipment operation information collected from sensors into process monitoring information and predicts plant equipment health according to the analysis results through a deep learning algorithm; and a manager terminal that receives the health prediction results of plant equipment from the failure prediction server. Includes. Preferably, a failure prediction server; is a data collection module that collects process monitoring information in real time using sensors installed inside the oil and gas plant; An anomaly detection module that detects abnormalities by analyzing collected process monitoring information; A deep learning module that implements a plant equipment failure prediction model using abnormal symptom detection results and process monitoring information as a learning data set; and a health prediction module that predicts the health of oil and gas plant equipment by inputting process monitoring information into a plant equipment failure prediction model; Includes.

In an embodiment, the failure prediction server preprocesses abnormality diagnosis data including plant equipment operation information for abnormality diagnosis, calculates the severity of outliers for each tag based on the difference between the abnormality score predicted by the failure prediction model and the standard, and calculates. Outlier severity for each tag is sorted in ascending or descending order, and inspection tags are recommended based on the tags with the highest severity.

The failure prediction diagnosis system and method through pattern analysis according to the failure type as described above selects and applies an appropriate model according to the state of the data of the oil plant facility and determines the timing of predictive maintenance by diagnosing the state or estimating the failure rate. By doing so, it is possible to operate oil plant facilities stably.

In addition, through examples, the failure prediction rate of oil plant facilities can be improved by applying a machine learning prediction model that applies unstructured data to overcome the limitations of the low failure prediction rate caused by the underutilization of structured and unstructured data in existing statistical techniques.

The effects of the present invention are not limited to the effects described above, and should be understood to include all effects that can be inferred from the configuration of the invention described in the detailed description or claims of the present invention.

Figure 1a is a diagram showing the configuration of an oil and gas plant equipment failure prediction system according to an embodiment
FIG. 1B is a diagram illustrating the data processing process of a failure prediction diagnosis system through pattern analysis according to failure type according to an embodiment.
Figure 1C is a diagram showing the data processing process of the failure prediction server using abnormality diagnosis data according to an embodiment.
Figure 2 is a diagram showing the data processing configuration of a failure prediction server according to an embodiment
Figure 3 is a diagram showing process monitoring information collected from a sensor according to an embodiment
Figure 4 is a time series graph showing vibration data detected in the drain line and buffer gas line.
Figures 5 and 6 are diagrams for explaining the learning data set data construction process according to an embodiment
7 is a diagram for explaining a method for determining abnormality signs according to an embodiment
Figure 8a is a diagram for explaining a pattern recognition method compared to the DCS alarm system according to an embodiment
Figure 9 is a diagram showing the data processing flow for oil and gas plant equipment failure prediction according to an embodiment
10 is a diagram illustrating a user interface of a failure prediction system according to an embodiment.
11 is a diagram for explaining the configuration of an inspection site recommendation screen according to an embodiment

The advantages and features of the present invention and methods for achieving them will become clear by referring to the embodiments described in detail below along with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below and may be implemented in various different forms. The present embodiments are merely provided to ensure that the disclosure of the present invention is complete and to provide common knowledge in the technical field to which the present invention pertains. It is provided to fully inform those who have the scope of the invention, and the present invention is only defined by the scope of the claims. Like reference numerals refer to like elements throughout the specification.

In describing embodiments of the present invention, if it is determined that a detailed description of a known function or configuration may unnecessarily obscure the gist of the present invention, the detailed description will be omitted. The terms described below are terms defined in consideration of functions in embodiments of the present invention, and may vary depending on the intention or custom of the user or operator. Therefore, the definition should be made based on the contents throughout this specification.

Figure 1a is a diagram showing the configuration of an oil and gas plant equipment failure prediction system according to an embodiment.

Referring to FIG. 1A, an oil and gas plant equipment failure prediction system according to an embodiment may be configured to include a sensor 100, a failure prediction server 200, and an administrator terminal 300. In an embodiment, the sensor 100 is installed in an oil and gas plant, senses plant equipment operation information, and transmits it to the server. The failure prediction server 200 collects and analyzes plant equipment operation information collected from sensors as process monitoring information, and predicts plant equipment health according to the analysis results through a deep learning algorithm. The manager terminal 300 receives the health prediction results of plant equipment from the failure prediction server and performs a series of procedures such as inspection and replacement based on the health prediction results.

The failure prediction diagnosis system and method through pattern analysis according to the failure type according to the embodiment provides a failure prediction diagnosis system through pattern analysis according to the failure type of oil and gas plant core equipment.

In the embodiment, data is analyzed using machine learning and deep learning for failure prediction and diagnosis, and the failure rate is estimated through an artificial neural network model to diagnose device status or failure to predict the time of failure. We propose a foresight model. The diagnosis model can diagnose conditions or faults using a supervised learning SVM (Support Vector Machine)-based diagnosis model and an unsupervised learning One-Class SVM-based diagnosis model.

The health prediction model provided in the embodiment includes a Markov Chain Monte Carlo (MCMC)-based prediction model that analyzes Time To Event (TTE) data in a non-parametric method and a Long Short-Term Memory (LSTM) model that analyzes data in a parametric method. Using a predictive model, the failure rate over time can be estimated and the time of failure can be predicted by measuring the remaining useful life (RUL) before failure occurs.

In addition, the diagnostic model and health prediction model provided in the embodiment allow the selection and application of an appropriate model according to the state of the data, and the timing of predictive maintenance can be determined by diagnosing the state or estimating the failure rate.

In an embodiment, a deep learning neural network including at least one of a Deep Neural Network (DNN), a Convolutional Neural Network (CNN), a Recurrent Neural Network (RNN), a Bidirectional Recurrent Deep Neural Network (BRDNN), and a Bayesian network. A failure prediction model is implemented by learning with a training data set for failure diagnosis.

FIG. 1B is a diagram illustrating a data processing process of a failure prediction diagnosis system through pattern analysis according to failure type according to an embodiment.

Referring to FIG. 1B, in the embodiment, the failure prediction server 200 collects equipment data as learning data for a deep learning model for failure prediction and performs preprocessing such as labeling, removing missing values, and dividing abnormal sections. Afterwards, the preprocessed equipment purification data can be used as learning data, or the equipment purification data can be input into the deep learning model to drive the deep learning model. In an embodiment, the failure prediction server 200 learns a deep learning model using equipment purification data, stores the model, executes the stored model, and evaluates the execution results. Afterwards, the failure prediction server 200 visualizes the model execution and evaluation results and stores the evaluation results. In an embodiment, the deep learning model evaluation results may be stored as abnormal diagnosis data. In an embodiment, anomaly diagnosis data may be constructed based on the outlier score results returned from the anomaly diagnosis model and used as input to an anomaly prediction model. In an embodiment, the failure prediction server 200 may perform anomaly detection based on an unsupervised learning model.

Figure 1c is a diagram showing a data processing process of a failure prediction server using abnormality diagnosis data according to an embodiment.

Referring to FIG. 1C, in the embodiment, the failure prediction server 200 preprocesses the abnormality diagnosis data and calculates the severity of the abnormality for each tag. In the embodiment, the outlier severity for each tag is a severity value calculated based on the difference between the anomaly score predicted by the model for each tag and the anomaly standard. In the embodiment, the anomaly score is a value indicating the degree of anomaly in the data. In the embodiment, the outlier score is used as an indicator to measure the degree to which a given data point deviates from a general pattern. The anomaly score can be calculated using a machine learning-based anomaly detection model. In an embodiment, the abnormality standard may consist of alarm standard data predicted by the model and data containing actual alarm ringing values. Then, the outlier severity for each tag is sorted in ascending or descending order, and the tag with the highest severity is selected. In an embodiment, the failure prediction server 200 may select the top 5 or 10 tags. Afterwards, the failure prediction server 200 recommends an inspection tag based on the tag with the highest severity. Thereafter, the failure prediction server 200 calculates the performance of the components included in the recommended inspection tag. In an embodiment, precision and recall between the recommended tag and the tag that actually alerted can be calculated as performance.

In the embodiment, the failure prediction server 200 preprocesses data for abnormality diagnosis. Data for abnormality diagnosis includes information related to the operation of plant equipment, and preprocessing includes data purification, conversion, and scaling. Afterwards, the failure prediction server 200 calculates an abnormality score for each tag using a failure prediction model using the preprocessed data. The abnormality score is a number that indicates how much the plant equipment status of the corresponding tag deviates from normal. Afterwards, the failure prediction server 200

Compare the ideal score with the standard value and calculate the difference between the ideal score and the standard. The reference value is generally a value that represents the normal range. For example, if the reference value is 0 and the abnormality score is a positive number, it means that the tag is outside the normal range. Afterwards, the failure prediction server 200 calculates the severity of the outlier for each tag based on the difference between the anomaly score and the standard. In an embodiment, the severity of the calculated abnormal score may be evaluated according to the range that includes the difference between the abnormal score and the standard. The greater the difference between the abnormal score and the standard value, the higher the severity can be judged. The failure prediction server 200 calculates the severity of outliers for each tag and sorts them in ascending or descending order. This sorts the tags in order of increasing severity. In the embodiment, the failure prediction server 200 places tags with higher severity at the top and proceeds with the process of recommending inspection tags based on them. Through this, the embodiment compares the abnormality score obtained from the failure prediction model with the reference value to evaluate the severity of the abnormal value for each tag, and performs maintenance or inspection preferentially based on this.

Figure 2 is a diagram showing the data processing configuration of a failure prediction server according to an embodiment.

Referring to FIG. 2, the eunuch prediction server 200 according to the embodiment includes a data collection module 210, an abnormality detection module 220, a deep learning module 230, a health prediction module 240, and a display module 250. ) and a feedback module 260. The term 'module' used in this specification should be interpreted to include software, hardware, or a combination thereof, depending on the context in which the term is used. For example, software may be machine language, firmware, embedded code, and application software. As another example, hardware may be a circuit, processor, computer, integrated circuit, integrated circuit core, sensor, Micro-Electro-Mechanical System (MEMS), passive device, or a combination thereof.

The data collection module 210 collects process monitoring information in real time using sensors installed inside the oil and gas plant. In the embodiment, the sensor includes a temperature, pressure measurement sensor, a bearing temperature, and a vibration measurement sensor, and the process monitoring information collected from the sensor includes the operating pressure (Operating Pressure) and temperature (Temp) of each facility, as shown in FIG. 3. , Alarm Shutdown Point data, etc. Additionally, the data collected in the embodiment may include big data related to detecting the health of the target facility, failure history data, maintenance history data, operation data, etc. In addition, reliability information such as failure frequency, failure time, failure type, and cause of failure, maintenance information including maintenance person, time required, etc., and information on the status of each equipment including temperature, vibration, noise, and current, etc. I can hear it.

Additionally, the data collection module 210 collects learning data for a deep learning model used in a failure prediction diagnosis system. For example, the data collection module 210 collects learning data of deep learning models including a failure prediction model, health diagnosis model, and anomaly detection model.

The preprocessing module 215 preprocesses the collected learning data to remove data with bias or discrimination among the collected artificial intelligence learning data. In the embodiment, the preprocessor 120 preprocesses the collected data and processes it into a form suitable for learning an artificial intelligence model. For example, the preprocessor 120 may perform processes such as noise removal, outlier removal, and missing value processing. Additionally, the preprocessing module 215 can prevent the model from learning unnecessary patterns by normalizing data, removing outliers, or scaling data through data preprocessing.

The abnormality detection module 220 analyzes the collected process monitoring information and detects abnormalities. In the embodiment, the anomaly detection module 220 uses a sensor-based technique to detect process anomalies early, as shown in a of FIG. 7, when the sensor measurement value is greater than the threshold or when the change pattern is detected. When it differs from normal, it is judged as a process abnormality.

In addition, in the embodiment, the abnormality detection module 220 performs parameter analysis (as shown in b of FIG. 7) to determine a process abnormality when the difference between the parameters calculated with normal data and the parameters calculated with actual data is large. Through the Parameter Analysis (Parameter Analysis) method, abnormal signs can be identified. In addition, as shown in c of Figure 7, if the difference between the value predicted by the pattern recognition model made from normal data and the value measured by the actual sensor is large, the abnormality is detected through the residual analysis method that determines the process abnormality. You can recognize the signs.

In an embodiment, the anomaly detection module 220 may select a method for determining an anomaly depending on the type of device that needs to detect the anomaly and the type of data that needs to be analyzed. For example, the anomaly detection module 220 uses sensor-based techniques to detect abnormalities such as rotating machine vibration, sensor sticking, peak, or flame state, pump head, process efficiency, heat exchanger fouling, and reaction speed. Parameter analysis methods are used to determine abnormal signs such as

To this end, in the embodiment, the anomaly detection module 220 collects data related to process efficiency, heat exchanger fouling, reaction speed, etc. This records various parameter values for the operation of the system or process. Afterwards, the collected data is cleaned, missing values are processed, and converted to the required form. Through data preprocessing, the quality of data is improved and converted into a form suitable for analysis. Thereafter, in the embodiment, the anomaly detection module 220 visualizes the data and confirms the distribution and pattern of the data using a time series graph, histogram, box plot, etc. In addition, average, variance, correlation, etc. are analyzed through basic statistical analysis. Thereafter, in the embodiment, the anomaly detection module 220 estimates parameters through data and models factors affecting the system or process. For example, use linear regression models, time series analysis models, or machine learning models to explore relationships between data and parameters. Thereafter, in the embodiment, the anomaly detection module 220 analyzes anomalies occurring in the system or process based on the estimated parameters. For example, it detects pump head abnormalities, heat exchanger fouling, and changes in reaction speed.

In addition, the residual analysis method can be used to identify abnormal signs of sensors, facilities, and processes as failures included in the pattern recognition model. Figure 4 shows a time series graph of vibration data detected in the drain line and buffer gas line. In the embodiment, abnormal points in the vibration data can be identified through pattern and trend analysis of the graph.

In the embodiment, for residual analysis, an appropriate model is selected based on data from sensors, facilities, processes, etc. Select and learn the most appropriate model among various models such as linear regression, nonlinear regression, and time series analysis. Afterwards, the predictive power of the learned model is evaluated. In embodiments, performance indicators of the model, such as mean squared error (MSE) or coefficient of determination (R-squared), are used to evaluate how well the model explains the data. Afterwards, the anomaly detection module 220 calculates the residual after the model is learned. The residual is the difference between the actual observed value and the value predicted by the model (residual = actual observed value - model predicted value). Afterwards, the anomaly detection module 220 visualizes and analyzes the calculated residual. In the examples, emphasis is placed on confirming the normality, homoscedasticity, and independence of the residuals. Check whether the residuals follow a normal distribution, determine whether the residuals have constant variance, and whether the residuals are independent of each other. Afterwards, the anomaly detection module 220 detects outliers through residual analysis. Values with large residuals or residuals with inconsistent patterns are judged to have a high possibility of being outliers. Afterwards, the anomaly detection module 220 improves the model according to the outliers or model suitability found through residual analysis. Remove outliers or try a different model to make more accurate predictions.

In the embodiment, sensor-based techniques, parameter analysis methods, and residual analysis methods are used to detect abnormalities in order to select the optimal method according to the characteristics and analysis purpose of the collected data.

Sensor-based techniques are used for anomalies that deal with time series or real-time data, such as rotating machine vibration, sensor failure, and spark conditions. This is because data can be collected and analyzed in real time to quickly detect abnormalities. For example, a vibration sensor can be used to detect unusually high vibrations in a machine, or a flame sensor can be used to detect unusual flame conditions.

The parametric analysis method is used to detect abnormalities that depend on several parameters such as pump head, process efficiency, heat exchanger fouling, reaction speed, etc. The parametric analysis method trains a model using given data, changes in the corresponding parameters, and This is to analyze the relationship with the abnormal state. In other words, when it is necessary to detect anomalies in data that depend on multiple parameters, a parameter analysis method is used. Through this, it is possible to determine whether changes in specific parameters are associated with abnormalities.

The residual analysis method is used to identify and solve problems that occur in data from sensors, facilities, processes, etc. Residual analysis identifies abnormalities in the data by analyzing the residual, which is the difference between the model's predicted value and the actual data. This is to evaluate the suitability of models used in equipment and processes. This allows you to identify outliers that occur in the data and resolve them.

In addition, in the embodiment, the anomaly detection module 220 issues an alarm when an operating variable exceeds the boundaries of the first threshold (high value) and the second threshold (low value) in the DCS (Distributed Control System), which is a plant operation computer. Signs of abnormalities can be identified through pattern recognition techniques. Figure 8 is a diagram for explaining a pattern recognition method compared to the DCS alarm system according to an embodiment. Generally, the alarm boundary range is set wide to reduce errors in judging normal as abnormal. In the embodiment, as shown in FIG. 8, when the alarm boundary range is set narrow, the time difference (Δ) between the existing method and the newly set alarm boundary range is This allows process abnormalities to be detected as early as possible.

The deep learning module 230 uses abnormal symptom detection results and process monitoring information as a learning data set to implement a plant equipment failure prediction model. In an embodiment, the deep learning module 230 uses at least one of a Deep Neural Network (DNN), a Convolutional Neural Network (CNN), a Recurrent Neural Network (RNN), a Bidirectional Recurrent Deep Neural Network (BRDNN), and a Bayesian network. A health prediction model can be implemented by learning a deep learning neural network containing a training data set containing process monitoring information and anomaly detection information.

Figure 5 is a diagram for explaining a learning data set data construction process according to an embodiment. In an embodiment, the results of the data graph analysis shown in FIG. 5 may be written as text, and the written text may be converted into a learning data set through a natural language processing process. For example, the manager terminal has a history of vibration abnormalities that occurred within a week after facility inspection and maintenance, and abnormal signs of vibration and RTD temperature were observed 2-3 days ago. In the case of 2019.09.05, symptoms of vibration and Comp equipment stoppage, When you input text about equipment monitoring results, such as 'The abnormality sign was a decrease in RTD temperature and vibration from around 14:00 on 2019.09.03,' in the embodiment, the deep learning module recognizes the text through natural language processing and extracts keywords. , a learning data set can be constructed. Figure 6 is a diagram showing an example of a learning data set related to equipment history constructed in an embodiment. In the embodiment, as shown in FIG. 6, the 101 shared equipment history data can be organized by issue, analyzed in normal and abnormal states, and constructed as a learning data set.

In an embodiment, the deep learning module 230 monitors the system in real time to perform Prognostics and Health Management (PHM), detects abnormal signs early, predicts when a failure occurs, and suggests the necessary maintenance time.

Machine learning and deep learning used in the embodiment are techniques for learning using training data and verification data, extracting feature values, calculating weights, classifying, or analyzing correlations between data. In the embodiment, a model for diagnosing status or failure and estimating the failure rate is proposed by applying machine learning and deep learning to the diagnosis and prediction stages of PHM, and in the decision support stage, the estimated failure rate and predicted useful time (RUL) are applied. We propose a method to determine the timing of predictive maintenance using Life). The data used can be identification information and sensor data from C-MAPSS (Commercial Modular Aero-Propulsion System Simulation). Additionally, in the embodiment, in order to assume the state of the data to be analyzed, the One-Class SVM-based diagnosis model proceeds with the process of generating failure data by adding outliers to C-MAPSS. In the MCMC-based prognostic model, the process of converting C-MAPSS into TTE (Time To Event) is performed, which consists of time value and event value. The SVM (Support Vector Machine)-based diagnosis model is applied when normal data and failure data are classified to diagnose the condition, and the One-Class SVM-based diagnosis model is applied when normal data and failure data are not classified. Apply to diagnose failure.

The MCMC-based predictive model converts the data to TTE when there is insufficient data to be analyzed or has uncertainty, estimates the failure rate using simulation techniques, and configures the data in the form of TTF (Time To Failure) when there is sufficient data to be analyzed. The failure rate is estimated by applying a prediction model based on LSTM (Long Short-Term Memory). And in the decision support stage, we propose a method to measure RUL and determine the timing of predictive maintenance using the failure rate estimated by MCMC (Markov Chain Monte Carlo)-based predictive model and LSTM-based predictive model.

Additionally, in the embodiment, when the deep learning module 230 is in the early stages of applying a prognostic model or when there is not enough data to be analyzed, the data is converted to TTE and applied to an MCMC-based prognostic model using a non-parametric method to reduce the failure rate. Whether it can be estimated is verified through experiment. Conversely, if there is enough data to be analyzed, it can be verified through experiment whether the failure rate can be estimated through CDF (Cumulative Distribution Function) by applying it to the LSTM-based prediction model to create a weight model and calculating the Alpha and Beta of the Weibull distribution. . In addition, model performance evaluation indicators and DTW (Dynamic Time Warping) are used to verify model performance, and data preprocessing, machine learning, and deep learning used in the embodiments can be implemented using Python modules, packages, and libraries. Provides a system.

In an embodiment, the deep learning module 230 can extract a number of possible causes for plant abnormalities by probabilistically analyzing causal relationships between variables through a Bayesian network.

The health prediction module 240 predicts the health of oil and gas plant equipment by inputting process monitoring information into a plant equipment failure prediction model. Health prediction is the detection of damage within a hazardous structure in order to continuously monitor the structural status (health) and performance of the structure or component in operation. At this time, health can be defined as maintaining the original state, function, and performance of a structure or facility throughout its entire lifespan. The health prediction module 240 diagnoses and predicts the health of the equipment and determines whether there is damage including failure of materials or structures or functional failure.

The display module 250 matches and outputs visual effects and objects to the data health prediction range and abnormality detection results. In an embodiment, the display module 250 matches events whose health is below a certain level and events requiring prompt action for identified abnormalities with preset visual objects so that the manager can identify them first. For example, the display module 250 may output actual data and abnormal state prediction scores for the actual data in a graph.

Figure 8b is a diagram showing state prediction results for actual data according to an embodiment. As shown in FIG. 8B, in the embodiment, the state can be predicted using the outlier severity and performance calculation values for each tag.

Additionally, the display module 250 may output a user interface for failure prediction.

Figure 10 is a diagram showing a user interface of a failure prediction system according to an embodiment.

Referring to FIG. 10, in the embodiment, when the data set to be predicted is selected in the user interface and inputted into the deep learning model, the state prediction result is output as a graph. In the embodiment, as shown in FIG. 9, a tag recommending an inspection site may be displayed with a marker or the like.

Figure 11 is a diagram for explaining the configuration of an inspection site recommendation screen according to an embodiment.

Referring to FIG. 11, in the embodiment, the measured values and abnormality scores for each data type such as Temp, Pressure, and Vibration can be confirmed together on the inspection part recommendation screen.

In the embodiment, the top tag is selected based on the abnormality score, and inspection areas are recommended from the selected tag. In the embodiment, when failure prediction of the recommended inspection area is performed, abnormal cause analysis results are provided through the failure prediction model.

The feedback module 260 receives feedback on abnormal symptom judgment and health prediction results from the manager terminal, and updates the artificial neural network for health prediction and the algorithm used for abnormal symptom determination. In an embodiment, the feedback information provided by the manager terminal may include accuracy assessment of health prediction, actions for abnormal signs and events where health is below a certain level, and result information after action.

The feedback module 260 evaluates the learned artificial neural network model and deep learning model. In an embodiment, the feedback module 260 may evaluate the artificial neural network model through at least one of accuracy, precision, and recall. Accuracy is an indicator that measures how well the results predicted by an artificial neural network model match the actual results. Precision is an indicator that measures the proportion of results predicted to be positive that are actually positive. Recall rate is an indicator that measures the proportion of actual positives predicted by the model as positive. In an embodiment, the feedback module 260 may calculate the accuracy, precision, and recall rate of the artificial neural network model, and evaluate the artificial neural network model based on at least one of the calculated indicators.

In an embodiment, the feedback module 260 may measure the accuracy of the artificial neural network model using an evaluation dataset. The evaluation dataset consists of data that the model did not use for learning, and is used to objectively evaluate the model's performance. In an embodiment, the feedback module 260 executes an artificial neural network model using an evaluation dataset and compares the predicted value of the artificial neural network model for each input data with the actual correct value of the data. Afterwards, through the comparison results, you can measure how accurately the model predicts. For example, in the feedback module 260, accuracy may be calculated as the ratio of data correctly predicted by the model among all data.

In addition, the feedback module 260 calculates the F1 score, which is an indicator indicating the balance between precision and recall, which is an indicator calculated as the harmonic average of precision and recall, and evaluates the artificial neural network model based on the calculated F1 score. In addition, the AUC-ROC curve, which is an index that visualizes the performance of the classification model in a graph, can be generated, and the artificial neural network model can be evaluated based on the generated AUC-ROC curve. In an embodiment, the feedback module 260 may evaluate the model's performance as better as the area under the ROC curve (AUC) is closer to 1.

Additionally, the feedback module 260 may evaluate the interpretability of the artificial neural network model. In an embodiment, the feedback module 260 evaluates the interpretability of the artificial neural network model through SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) methods. SHAP (SHapley Additive exPlanations) is a library that provides interpretation of the results predicted by the model, and the feedback module 260 extracts SHAP values from the library. In an embodiment, the feedback module 260 can predict how much the characteristic information input to the model influenced the model prediction by extracting the SHAP value.

The LIME (Local Interpretable Model-agnostic Explanations) method is a method of explaining model predictions for individual samples. In an embodiment, the feedback module 260 calculates the importance of each characteristic information by approximating the sample to an interpretable model through the LIME method. Additionally, the feedback module 260 can estimate the influence of each characteristic variable by analyzing the internal weight and bias values of the model.

The feedback module 260 performs improvement work when the artificial neural network model shows low fairness or discrimination. In an embodiment, when the data for a specific group is insufficient by a certain level, the feedback module 260 additionally collects data representing the specific group and performs a data preprocessing process. In an embodiment, the feedback module 260 performs a data preprocessing process including data normalization, outlier removal, and data scaling to prevent the model from learning unnecessary patterns. Additionally, in an embodiment, the feedback module 260 may prevent discrimination or ensure fairness by adding a specific condition to the model learning algorithm.

In the embodiment, the feedback module 260 evaluates the performance of the model by comparing the predicted results of the model with the actual results through confusion matrix analysis to ensure fairness. The confusion matrix is a matrix that evaluates the classification performance of a model in supervised learning. The confusion matrix displays the classification results by comparing the results predicted by the model with the actual results. In an embodiment, the feedback module 260 may evaluate the performance of the model by calculating the accuracy and misclassification rate for each class through confusion matrix analysis.

Additionally, in the embodiment, the feedback module 260 allows the distribution of data to be confirmed through visual analysis of the learning data. For example, in the case of image data, the diversity and fairness of the data can be evaluated by visualizing image samples for each class.

Additionally, the feedback module 260 performs bias verification of the artificial neural network model. In the embodiment, the feedback module 260 verifies whether the model is biased toward a specific class or attribute through bias verification of the learning information. To this end, the feedback module 260 compares the number of samples for each class or evaluates classification performance for each class.

In addition, the feedback module 260 verifies the fairness and diversity of the learning data and improves the artificial neural network model through fairness verification and evaluation index calculation. In the embodiment, fairness verification is to check whether the artificial neural network model shows discrimination with respect to specific attributes such as gender, race, and age with respect to the learning information. In an embodiment, the feedback module 260 may check whether a specific attribute is differentiated by comparing the number of samples for each attribute or evaluating classification performance for each attribute.

Additionally, the feedback module 260 calculates various indicators to evaluate the performance of the artificial neural network model. For example, you can evaluate model performance by calculating indicators such as accuracy, precision, recall, and F1-score. At this time, the fairness and diversity of the model can be evaluated by calculating indicators for each class.

In addition, the feedback module 260 collects feedback on problems that occur when the artificial neural network model is used in a real environment, and continuously improves the artificial neural network model by reflecting the collected feedback into the artificial neural network model.

Below, methods for predicting oil and gas plant equipment failures will be explained in turn. Since the operation (function) of the oil and gas plant equipment failure prediction method according to the embodiment is essentially the same as the function of the counterfeit product detection system, descriptions overlapping with FIGS. 1 to 8 will be omitted.

Figure 9 is a diagram showing a data processing flow for predicting oil and gas plant equipment failure according to an embodiment.

Referring to FIG. 9, in step S100, the data collection module collects data generated during operation from sensors in plant equipment. In step S200, in order to facilitate analysis of the data collected from the data collection module, the data is converted to extract feature values or convert the size. In step S300, the equipment status is evaluated using data converted from the abnormal symptom judgment module. In the S300 stage, in order to detect signs of process abnormality early, a process abnormality is determined when the sensor measurement value is greater than the threshold or the change pattern is different from normal. Using sensor-based techniques, parameters calculated from normal data and actual data are used. Parameter analysis method, which determines a process abnormality if the difference between calculated parameters is large, and residual analysis, which determines a process abnormality if the difference between the value predicted by a pattern recognition model made from normal data and the value measured by the actual sensor is large. (Residual Analysis) method can be used to evaluate the condition of equipment and identify abnormal signs. In step S400, the health prediction module determines whether the condition has deteriorated and diagnoses the failure. In the S400 stage, the deep learning module probabilistically analyzes causal relationships between variables through a Bayesian network to extract a number of possible causes for plant abnormalities. In the S500 stage, the health prediction module identifies failures through effective time (RUL). Predict when it will happen. In an embodiment, step S500 may include a decision support step that suggests a time for predictive maintenance before a failure occurs, and a human-machine interface step that informs the user of data generated in previous steps through an interface such as a dashboard. You can.

The failure prediction diagnosis system and method through pattern analysis according to the failure type as described above selects and applies an appropriate model according to the state of the data of the oil plant facility and determines the timing of predictive maintenance by diagnosing the state or estimating the failure rate. By doing so, it is possible to operate oil plant facilities stably.

In addition, through examples, the failure prediction rate of oil plant facilities can be improved by applying a machine learning prediction model that applies unstructured data to overcome the limitations of the low failure prediction rate caused by the underutilization of structured and unstructured data in existing statistical techniques.

The disclosed content is merely an example, and various modifications and implementations may be made by those skilled in the art without departing from the gist of the claims, so the scope of protection of the disclosed content is limited to the above-mentioned specific scope. It is not limited to the examples.

Claims (10)

In the oil and gas plant equipment failure prediction system,
Sensors installed inside oil and gas plants to sense plant equipment operation information;
A failure prediction server that collects and analyzes plant equipment operation information collected from the sensor into process monitoring information and predicts plant equipment health according to the analysis results through a deep learning algorithm; and
A manager terminal receiving a health prediction result of plant equipment from the failure prediction server; Including,
the failure prediction server; Is
A data collection module that collects process monitoring information in real time using sensors installed inside oil and gas plants;
An abnormality detection module that detects abnormalities by analyzing the collected process monitoring information;
A deep learning module that implements a plant equipment failure prediction model using abnormal symptom detection results and process monitoring information as a learning data set;
a health prediction module that predicts the health of oil and gas plant equipment by inputting process monitoring information into the plant equipment failure prediction model; and
A display module that matches and outputs visual effects and objects to the data health prediction range and abnormality detection results; contains
the failure prediction server; Is
For abnormality diagnosis, abnormality diagnosis data including plant equipment operation information is pre-processed, and the severity of outliers for each tag is calculated based on the difference between the abnormality score predicted by the failure prediction model and the standard. The severity of outliers for each tag is calculated in ascending order. Or sort in descending order and recommend inspection tags based on the highest severity tags,
The abnormality detection module; silver
In order to detect signs of process abnormalities early, process abnormalities are determined when sensor measurements are greater than the threshold or change patterns are different from normal. Sensor-based techniques, parameters calculated from normal data and parameters calculated from actual data Parameter Analysis, which determines a process abnormality if the difference is large, and Residual Analysis, which determines a process abnormality if the difference between the value predicted by a pattern recognition model made from normal data and the value measured by the actual sensor is large. ) method to identify abnormal signs and
The abnormality detection module; silver
DCS (Distributed Control System), a plant operation computer, identifies abnormal signs through a pattern recognition technique that issues an alarm when operating variables exceed the boundaries of the first and second thresholds.
The deep learning module; silver
By probabilistically analyzing causal relationships between variables through a Bayesian network, a number of possible causes for plant abnormalities are extracted.
The abnormality detection module; silver
Evaluate models through model performance indicators, including mean squared error (MSE) or coefficient of determination (R-squared), and select a model suitable for residual analysis based on data from sensors, facilities, and processes. do,
After learning the selected model, calculate the residual, which is the value obtained by subtracting the actual observed value from the value predicted by the model, and visualize the calculated residual.
In order to check the normality, homoscedasticity, and independence of the residuals, check whether the midpoint residuals follow a normal distribution, check whether the residuals have a certain variance, and determine whether the residuals are independent of each other.
The abnormality detection module; silver
Using the results of residual analysis, values with residuals above a certain level or residuals with inconsistent patterns are judged as residuals with the possibility of being outliers.
The abnormality detection module; silver
DCS (Distributed Control System), a plant operation computer, identifies abnormal signs through a pattern recognition technique that issues an alarm when operating variables exceed the boundaries of the first threshold (high value) and second threshold (low value).
The abnormality detection module; silver
The failure rate and effective time are estimated using the identification information of C-MAPSS (Commercial Modular Aero-Propulsion System Simulation), sensor data, and artificial intelligence model, and the estimated failure rate and predicted effective time (RUL, Remaining Useful Life) are used to estimate the failure rate and useful life. Decide on maintenance time,
The abnormality detection module; silver
In order to assume the state of the data to be analyzed, failure data is generated by adding outliers to C-MAPSS through a One-Class SVM-based diagnostic model.
The sensor is
It includes temperature and pressure measurement sensors, bearing temperature, and vibration measurement sensors, and the process monitoring information collected from the sensors includes operating pressure, temperature, and alarm shutdown point data of each facility. A plant equipment failure prediction system characterized by:
delete delete The method of claim 1, further comprising: the deep learning module; silver
Process monitoring information for deep learning neural networks including at least one of DNN (Deep Neural Network), CNN (Convolutional Neural Network), RNN (Recurrent Neural Network), BRDNN (Bidirectional Recurrent Deep Neural Network), and Bayesian network. and a plant equipment failure prediction system that implements a health prediction model by learning with a training data set containing abnormal symptom detection information.



delete delete delete delete delete delete