KR20230125116A - The method for fault detection without training data or diagnosis with prediction of remaining time until breakdown using clustering algorithm and statistical methods - Google Patents

Abstract

Its purpose is to provide a system and method for detecting defects in mechanical equipment, diagnosing failures, and predicting remaining time using clustering algorithms and statistical techniques of machine learning. In particular, quantitatively detect defects in the absence of data obtained from a system in a faulty state, such as a system at the beginning of operation, or diagnose a fault from a fault if there is system data in the faulty state to predict the time until failure aims to

Description

The method for fault detection without training data or diagnosis with prediction of remaining time until breakdown using clustering algorithm and statistical methods}

The present invention relates to a method for detecting faults, diagnosing and predicting faults in a system including mechanical facilities and devices based on machine learning and statistical techniques, and more particularly, in the absence of fault data through real-time data collection of the system. It relates to a system and method for detecting a fault in a system, diagnosing a fault to occur from a fault in the system, and predicting a time to fault.

Currently, in the management of most mechanical facilities and systems, system experts have diagnosed defects and failures and performed repairs or maintenance. However, it is difficult for facility experts to respond to invisible defects and failures, and they rely on regular preventive maintenance performed at regular intervals. Regular maintenance also results in cost loss as unnecessary parts are replaced regardless of actual defects.

In order to solve this problem, Korean Patent Publication No. 10-2019-0065718 and Korean Patent Publication No. 10-2019-0059561 disclose a method for diagnosing and predicting a failure by learning system failure data with an artificial intelligence model. Briefly, the failure data of the system is pre-learned in an artificial intelligence model, and the failure of the system is diagnosed from real-time sensing data using the artificial intelligence model. In this case, the learned artificial intelligence model receives real-time sensing data and determines the type of failure.

However, the prior art uses arbitrarily generated abnormal data to learn an artificial intelligence model as failure data of the system, which may differ from data that can be extracted from the actual system, and the failure of the system in visual form. It is difficult to make a quantitative judgment by diagnosing It also determines system anomalies, but does not predict the remaining time from an anomaly to a diagnosed failure.

Patent Publication 10-2020-0139346 Patent Publication 10-2020-0118743

An object of the present invention is to provide a system and method for detecting defects in mechanical equipment, diagnosing failures, and predicting remaining time using a clustering algorithm and statistical techniques of machine learning. In particular, quantitatively detect defects in the absence of data obtained from a system in a faulty state, such as a system at the beginning of operation, or diagnose a fault from a fault if there is system data in the faulty state to predict the time until failure aims to

The problem to be solved by the present invention is not limited to the problems mentioned above, and will be clearly understood by those skilled in the art from the following description.

In order to solve the above technical problem, a fault detection and fault diagnosis prediction method for a system without fault data using a machine learning clustering algorithm and statistical techniques according to the present invention is a temperature collected from a main sensor attached to a system such as a mechanical facility. a data receiving and storing unit for receiving and storing data such as , humidity, and vibration and all data related to failures that can be directly extracted from the inside of the system; extracting a feature from a dataset or real-time data in a normal or failed state input from the data reception and storage unit; Clustering the extracted features using a clustering algorithm as an unsupervised learning method of machine learning; defining a classification index of the specified clusters as a state index; In the case of the entire dataset in a normal or faulty state, defining a parameter of a probability distribution according to the state of the system; calculating a confidence interval from the probability distribution of the state indicator corresponding to the defined normal state. a defect detection unit for detecting a defect together with a status indicator of real-time data; It includes a failure diagnosis and prediction unit that calculates a confidence interval from the probability distribution of the condition indicators corresponding to the defined failure conditions, diagnoses failures together with the condition indicators of real-time data, and predicts the remaining time from the time the failure occurs to the failure. .

According to one embodiment of the present specification, the data reception and storage unit may include a dataset extracted in a normal state of the system and a dataset extracted in various failure states of the system.

According to one embodiment of the present specification, the data reception and storage unit collects the data received in real time in the normal state of the system to create a dataset in the normal state and use it as the dataset in the normal state.

According to an embodiment of the present specification, after calculating a condition indicator confidence interval from data obtained from a system in a normal state, real-time data may be received from the system and a condition indicator may be calculated to detect a defect.

According to an embodiment of the present specification, when there is system data in a normal state and a failure state, a condition indicator confidence interval is calculated from data obtained from the system in a normal state and a failure state, and then real-time data is received from the system and the condition index It is possible to diagnose and predict failures by calculating .

According to one aspect of the present specification, it is possible to quantitatively determine a system defect only with data in a normal state, away from the need for data in a normal state or a failure state for learning existing data. In particular, it is possible to more accurately detect system defects only with directly acquired data, breaking away from the limitations of learning artificial intelligence models by randomly creating outliers in the conventional data.

According to another aspect of the present specification, if there is data acquired from a system in a faulty state, it is possible to predict the time required to fail using a change in a condition indicator as well as fault diagnosis. In particular, it is possible to predict the time when the diagnosed failure will occur from the time when the defect is detected, out of the limitation of determining only the type of failure of the conventional artificial intelligence model.

The effects of the present invention are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the description below.

1 is a block diagram showing the configuration of a fault detection and fault diagnosis prediction method for a system without fault data using a clustering algorithm and a statistical technique according to an embodiment of the present invention.
2 is a flowchart of a fault detection and fault diagnosis prediction method for a system without fault data using a clustering algorithm and a statistical technique according to an embodiment of the present specification.
3 is a flowchart of constructing a state indicator by receiving data from the model learning unit 120 according to an embodiment of the present invention.
FIG. 4 is an exemplary view of deriving a state indicator value by appropriately dividing data received by the model learning unit 120. Referring to FIG.
5 is a flowchart of a defect detection method according to an embodiment of the present specification.
6 is an exemplary graph showing the distribution of state indicators in a steady state.
7 is an exemplary graph of state indicator values calculated by receiving data received in real time from the data reception and storage unit 110 and a 95% confidence interval of a state indicator in a steady state into the model learning unit 120 .
8 is a flowchart of a failure diagnosis and prediction method according to an embodiment of the present specification.
9 is a graph illustrating specific breakdown A (Breakdown A) and breakdown B (Breakdown B) state indicator confidence intervals among the various failure states, respectively.
10 is an example of a graph modeled by a primary equation in the failure diagnosis and prediction unit 140 after receiving real-time sensor data from the data reception and storage unit 110 into the model learning unit 120 and calculating a state index value.
11 is a graph example of a change trend of a real-time condition indicator value toward a fault B as opposed to a change trend of the real-time condition indicator toward the fault A.
12 is an example of a graph modeled with a quadratic curve for real-time state index values.

The features and advantages of the invention disclosed herein, and how to achieve them, will become clear with reference to the detailed description of the following embodiments taken in conjunction with the accompanying drawings. However, the present specification is not limited to the embodiments disclosed below and may be implemented in a variety of different forms, only the present embodiments are intended to complete the disclosure of the present invention, and are common in the art to which this specification belongs. It is provided to fully inform those skilled in the art of the scope of the invention, and the scope of rights in this specification is only defined by the scope of the claims.

In addition, in order to efficiently explain the technical components constituting the present invention, a preferred embodiment of the present invention to be carried out below is provided in each system function configuration or a system function configuration commonly provided in the technical field to which the present invention belongs It is omitted as much as possible, and the functional configuration that should be additionally provided for the present invention will be mainly described. If a person of ordinary skill in the art to which the present invention pertains, he/she will be able to easily understand the functions of conventionally used components among the omitted functional configurations, and also include the omitted components and the added components for the present invention as described above. The relationships between the components will also be clearly understood.

Terms used in this specification are for describing the embodiments and are not intended to limit the scope of the present specification. In this specification, singular forms also include plural forms unless specifically stated otherwise in a phrase. In this specification, “comprising” and/or “comprising” as used herein does not exclude the presence or addition of one or more other elements other than the recited elements.

Like reference numerals throughout the specification refer to like elements and “and/or” includes each and every combination of one or more of the recited elements.

Unless otherwise defined, all terms (including technical and scientific terms) used in this specification may be used with meanings commonly understood by those skilled in the art to which this specification belongs. In addition, terms defined in commonly used dictionaries are not interpreted ideally or excessively unless explicitly specifically defined.

<Term definition>

Prior to the description of the fault detection and fault diagnosis prediction method for a system without fault data using the machine learning clustering algorithm and statistical techniques according to the present specification, some terms will be defined first. In this specification, the fault detection and fault diagnosis prediction method of a system without fault data using a machine learning clustering algorithm and statistical techniques is applied to a system such as a machine that can obtain data by attaching a sensor or receive data from the system. As a method that can be applied, a system such as an airplane will be assumed.

Data that can be transmitted by attaching a sensor such as system vibration, temperature, humidity, etc. is called 'sensor data', and all data related to the state of the system from a device such as a computer built into the system and the sensor data. are collectively referred to as 'system data'.

Characterizing the failure of any one of several failure mechanisms during system operation is called diagnosis. In addition, a situation in which the system stops while the system is operating is called a 'breakdown', and a situation in which the system operates but abnormal symptoms are known is called a 'fault'.

Numerical indicators that indicate the state of the system are called 'condition indicators'. In this specification, a normal state and various failure states are represented as a condition indicator distribution to detect defects and diagnose and predict failures. To explain in more detail, the system detects defects from real-time data by setting the range of condition indicator values that can be displayed in a normal state, and if there is data obtained in various failure states of the system, the range of condition indicator values in a fault state to diagnose failures from real-time data and predict the remaining time until failure.

In this specification, the condition indicator is used to determine a normal state and a failure state of the system. According to the present specification, the machine learning algorithm used herein to define the condition indicator from data is a clustering algorithm of unsupervised learning, and how well the clustering of data has been achieved through this algorithm. The numerical index to be evaluated is named 'clustering evaluation index'. According to the present specification, the clustering evaluation index is used as a condition index.

The characteristics of the data of the system to be learned by the machine learning algorithm are called 'features'. In this specification, statistical features, kurtosis, and skewness in the statistical sense of data are presented as examples to detect defects in systems without failure data using machine learning clustering algorithms and statistical techniques. And a failure diagnosis prediction method will be described.

Kurtosis is an indicator of how close the data is to the center. Also, skewness is an index that indicates the degree of asymmetry of the data distribution compared to the normal distribution. According to one embodiment of the present specification, the reason for using the statistical characteristics such as skewness and kurtosis is suitable for representing the state of a system that changes over time. As an example, when the system is in the same state, it should exist around the same value while including the error of the process of measuring the data. Therefore, in order to train a machine learning algorithm model, newly creating or adjusting a feature that can best explain data is called feature extraction and feature engineering.

According to one embodiment of the present specification, a process of dividing data from the system data is performed. At this time, when data of a certain size can be accumulated in real time or data of a certain size can be received from the system, this is called a 'dataset'. In addition, to apply statistical characteristics to the clustering algorithm, the entire data set is divided at regular time intervals and used. In this case, the divided data is called a sub dataset. According to the present specification, a feature extraction process is applied to the sub data set and one feature is extracted. In this process, features such as the number of sub-data sets are extracted. In this case, the extracted features are called a feature dataset. After applying the clustering algorithm to the following feature data set, one condition indicator is extracted. At this time, a set of several condition indicators for normal or faulty conditions is called an indicator data set. Several condition indicator values are needed to calculate the range of condition indicators for a system. Therefore, the entire data that can compose multiple data sets from which one condition indicator value can be extracted is called 'full dataset'. According to the present specification, when the distribution of condition indicators is not well formed, it is possible to adjust the interval of the data set.

The method for detecting defects in a system without failure data and predicting failure diagnosis using a machine learning clustering algorithm and statistical techniques according to the present specification can extract various statistical characteristics of data and apply them to a clustering algorithm. Therefore, we will briefly describe some of the key statistical features used in the examples herein.

<Extracting statistical features>

Skewness is an indicator of the asymmetry of the distribution of data. In statistical analysis, since normal distribution of data is assumed in many cases, skewness is used as an indicator to see how far away from this normal distribution is.

In Equation 1 above, ' ' represents the skewness value and ' ' is the set of data, ' is the mean of the data, ' ' is the standard deviation of the data, means average.

If the skewness value has a positive value, the center of the data is skewed to the left of the normal distribution, and if it has a negative value, it means that the center of the data is skewed to the right of the normal distribution.

Kurtosis is an indicator of how close the data are to the mean.

In Equation 2 above,' ' means the kurtosis value, and if the value is 3, it is close to a normal distribution, if it is less than 3, the data is widely distributed, and if it is greater than 3, it can be judged that the data is concentrated in the center.

<Clustering Algorithm>

The method for detecting faults and diagnosing faults in a system without fault data using a machine learning clustering algorithm and statistical techniques according to the present specification may use a clustering algorithm among unsupervised learning machine learning methods. Therefore, the K-means clustering algorithm used in the examples of this specification will be briefly described.

The meaning of Equation 3 above is that the clusters including the data do not have common data and the center of each cluster ' Form clusters by minimizing the sum of the distances of data from '.

The model learning unit 120 extracts one value as a scale representing the degree of classification between the classified clusters of each sub-data set. The reason for using the clustering evaluation index is that when data is classified into two clusters, normal and abnormal, if the classification is well done, it can be regarded as a failure state. This is because it can be judged that it is close to the normal state because it is not. Therefore, the Davies-Bouldin Index (DBI) will be briefly described as a clustering evaluation index used in the examples of this specification.

<Clustering evaluation index>

Equation 4

In Equation 4 above, ' ' represents the distance between clusters, and ' ' represents the distance between data within a cluster.

The method for detecting faults and diagnosing faults in a system without fault data using a machine learning clustering algorithm and statistical techniques according to the present specification learns normal or faulty data using DBI as a clustering evaluation index after clustering a dataset and learns fault data. can detect

The model learning unit 120 calculates a 95% confidence interval by modeling a sampling distribution suitable for the DBI values (i = 1 to n) of each extracted sub-dataset (i = 1 to n).

<95% confidence interval according to normal distribution modeling>

The fault detection and fault diagnosis prediction method of a system without fault data using the machine learning clustering algorithm and statistical techniques according to the present specification may use a 95% confidence interval based on statistical hypothesis testing. Parameters such as average and variance of system data can be estimated from data measured by sensors. However, system fault detection and fault diagnosis prediction can be made using the fact that the parameters estimated when the system state is changed are different from the parameters of the previous state. Therefore, we briefly describe the 95% confidence interval of the sampling distribution.

In general, statistical hypothesis testing can extract data from one group (one system with fixed parameters) or extract data from two or more groups and compare multiple groups from the comparison of estimated parameters. According to an embodiment of the present specification, a 95% confidence interval corresponding to a Z-test method can be used as a method for comparing differences in sample means.

In the case of the statistical model proposed in an embodiment of this specification, the Z-test statistic was used on the assumption that the data that can be extracted from the system follows a normal distribution, but a T-test or the like can be used depending on the assumption of the system data. can

Example

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. On the other hand, for convenience of understanding, it is assumed that the fault detection and fault diagnosis prediction method for a system without fault data using the machine learning clustering algorithm and statistical technique according to the present specification is implemented as a computer program and executed by a computer processor. Explain. Therefore, it can be understood that each step of the fault detection and fault diagnosis prediction method for a system without fault data using the machine learning clustering algorithm and statistical techniques according to the present specification is executed by a processor.

1 is a block diagram showing the configuration of a fault detection and fault diagnosis prediction method system for a system without fault data using a machine learning clustering algorithm and a statistical technique according to an embodiment of the present invention, and FIG. 2 is an embodiment of the present invention. It is a flow chart according to an example. As shown in FIG. 1, the subject includes a data receiving unit 110, a model learning unit 120, a defect detection unit 130, a failure diagnosis and prediction unit 140, and a display unit 150.

Referring to FIG. 2 , first, in step S10, the data receiving and storing unit 110 of FIG. 1 receives and stores data transmitted in real time from a sensor attached to a machine or facility or from inside the machine or facility.

Here, the machine or facility includes all mechanical devices normally included in factory facilities, airplanes, or automobiles. It includes not only devices that can transmit data internally, such as computers, but also all equipment to which sensors can be attached.

In addition, sensors that can be attached to machines or facilities include all types of sensors that can affect mechanical facilities, including vibration sensors, humidity sensors, temperature sensors, and acceleration sensors, and data reception and storage units are wired and wireless networks. Acquires and stores sensing data and internal data in real time using Here, the wired and wireless networks include any one of 3G, 4G, 5G, LTE, WiFi, Bluetooth, and Zigbee.

In the next step S20, the model learning unit 120 of FIG. 1 receives the data stored in the data reception and storage unit 110 and models an index for distinguishing a normal state or a failure state.

Here, the state modeling in step S20 also includes data modeling of the normal state in the absence of failure data and data modeling of the normal state and the failure state in the case of failure data.

3 is a flowchart illustrating a detailed method of step S20 of a method for detecting a fault in a system without fault data and predicting a fault diagnosis using a machine learning clustering algorithm and a statistical technique according to an embodiment of the present invention; is an example of one indicator created for one defect detection and failure diagnosis and prediction using sub datasets divided at regular intervals from the entire dataset.

Hereinafter, with reference to FIGS. 3 and 4, the process of learning the statistical model of the normal state and the failure state of step S20 will be described in detail.

Referring to FIG. 3 , the model learning unit 120 receives an entire dataset in a normal state or a failure state. Here, the entire dataset means sensor and internal data extracted from a system in a normal or broken state having a size sufficient for learning, not data input in real time.

Referring to FIG. 4, the entire dataset is divided at regular intervals to create N subdatasets (Subdataset_{i = 1 to n}), and statistical features are extracted for each of the divided subdatasets. For convenience of calculation, it is assumed that two statistical characteristics, skewness and kurtosis, are used.

Various statistical characteristics including skewness and kurtosis used in the examples herein may be used, and various known statistical characteristics may be used without being limited to the statistical characteristics described.

Referring back to FIGS. 3 and 4 , the model learning unit 120 classifies the statistical features into two clusters by applying a machine learning clustering algorithm to each sub-data set as data. do.

According to an embodiment of the present specification, the classified two clusters use K-means clustering among clustering models corresponding to unsupervised learning of machine learning in machine learning. The reason for classifying into two clusters in the clustering model is that the system data can represent both normal state data and failure state data, so good clustering means that the failure state is clearly observed.

Referring back to FIGS. 3 and 4 , the model learning unit 120 extracts one value as an index representing the degree of classification between classified clusters of each sub-data set. The reason for using the clustering classification degree index is that when data is classified into two clusters, normal and abnormal, if the classification is done well, it can be regarded as a failure state. This is because it can be judged that it is close to the normal state.

According to an embodiment of the present specification, the Davies Bouldin Index (DBI) may be used as the clustering evaluation index.

Hereinafter, with reference to FIGS. 5 and 6 , a defect detection process of the defect detection unit 130 using the 95% confidence interval of the steady state acquired by the model learning unit 120 will be described in detail.

5 is a flowchart illustrating a detailed method of step S30 of a method for detecting a fault in a system without fault data and predicting a fault diagnosis using a machine learning clustering algorithm and a statistical technique according to an embodiment of the present invention.

The defect detection unit 130 may receive a condition indicator calculated through the model learning unit 120 from a previously collected dataset in a normal state.

The defect detection unit 130 may estimate a parameter of the state indicator of the steady state received from the model learning unit 120 assuming an appropriate statistical distribution and calculate a 95% confidence interval.

6 is an example showing the distribution of the condition indicators.

Referring to FIG. 6 , it can be seen that the state index in the steady state shows a normal distribution and is concentrated at one value.

According to an embodiment of the present specification, the statistical distribution is assumed to be a normal distribution, and parameters of the normal distribution corresponding to the state index of the steady state may be calculated through the following equation.

In Equation 5 above, ' ' is transmitted from the model learning unit 120 ' 'Dog clustering evaluation index value' '. ' ' is the average of clustering evaluation indicators, ' ' represents the standard deviation.

The clustering evaluation index may be used as a state index of the system, and according to an embodiment of the present specification, the range of the state index in a steady state may be calculated using the following formula using the estimated parameter.

' ' is the statistic corresponding to the 95% significance level of the Z-test, ' ' is a state indicator calculated by receiving the data received in real time from the data reception and storage unit 110 into the model learning unit 120.

If the state of the system is constant, the state indicator value calculated from real-time system data ' ' must fall within the range of the steady-state confidence interval calculated in Equation 6 above. Therefore, the fault detection method of fault detection and fault diagnosis prediction method of a system without fault data using machine learning clustering algorithm and statistical technique calculates the 95% confidence interval of the condition indicator in the normal state, and the data reception and storage unit (110 ) is input to the model learning unit 120 in real time, and when the calculated state indicator is out of the range of the steady state state indicator confidence interval, it is detected as a defect by a statistical test method.

7 is an exemplary graph of state indicator values calculated by receiving data received in real time from the data reception and storage unit 110 and a 95% confidence interval of a state indicator in a steady state into the model learning unit 120 .

Referring to FIG. 7 , the 95% confidence interval (dotted line) of the state indicator in the steady state calculated from Equation 6 can be confirmed as (0.91193, 0.92389). And the state indicator value calculated by receiving the dataset having the same size as the sub-dataset received in real time from the data receiving and storing unit 110 into the model learning unit 120 [Time 1: 0.92133, Time 2: 0.9157] , time 3: 0.920455, time 4: 0.90521]. In chronological order, 0.92133, 0.9157, and 0.920455 are included in the 95% confidence interval of the state indicator of the steady state, so the system can be judged to be in a steady state. However, the condition index value corresponding to time 4, 0.90521, is outside the 95% confidence interval of the condition index in the steady state. That is, it means that a defect was detected at time 4 by comparing the calculated state indicator value inputted from the data reception and storage unit 110 to the model learning unit 120 in real time with the calculated steady state state indicator confidence interval.

Hereinafter, with reference to FIG. 7 , a defect detection process of the defect detection unit 130 using the 95% confidence interval of the normal state acquired by the model learning unit 120 will be described in detail.

7 is a diagram to explain a detailed method of fault diagnosis and prediction in step S40 of a fault detection and fault diagnosis prediction method for a system without fault data using a machine learning clustering algorithm and a statistical technique according to an embodiment of the present invention. It is a flow chart.

Referring to FIG. 7 , failure diagnosis and prediction section 140 diagnosis and prediction of a failure detection and failure diagnosis and prediction method for a system without failure data using a machine learning clustering algorithm and a statistical technique according to an embodiment of the present invention. Unlike the defect detection unit 130 that detects a defect only with data in a normal state, the method can diagnose a failure state by using both normal data and data in a failure state and estimate the time from failure to failure.

Referring again to FIG. 7 in detail, the defect detection unit 130 transfers defect detection information to the failure prediction and detection unit 140 as the condition indicator value received in real time deviate from the normal state condition indicator confidence interval. Then, the sensor data received from the data reception and storage unit 110 is stored in real time from the time of receiving the defect detection information from the defect detection unit 130 (the time when the first value of the condition indicator value set is generated). A curve that can best explain the trend of change in state indicator values calculated by receiving input from the model learning unit 120 is estimated. In addition, a fault for failure A or failure B may be diagnosed according to the direction in which the real-time state indicator values change over time. In addition, it is possible to calculate the point in time when a failure occurs using the estimation curve.

Like the defect detection unit 130, the failure diagnosis and prediction unit 140 inputs data in a normal state to the model learning unit 120 to calculate a condition indicator, and various failure state datasets are also included in the model learning unit ( 120) to calculate the condition indicator.

Confidence intervals may be obtained for the calculated distribution of condition indicators in the normal state and the distribution of condition indicators in various failure states.

According to an embodiment of the present specification, the statistical distribution is assumed to be a normal distribution, and parameters of the normal distribution corresponding to the condition indicator distribution in the normal state and the condition indicators in various failure states can be calculated through Equation 5 .

9 is a graph illustrating specific failure A and failure B state indicator confidence intervals among the various failure states, respectively.

Referring to FIG. 9, the distribution and 95% confidence interval for failure A and failure B, respectively. As can be seen in FIGS. 9 and 10, the average and variance of the state indicator of failure A are 0.651571 and 0.002246, and the 95% confidence interval through Equations 5 and 6 is (0.647169, 0.655972), and the average and variance of the state indicator of failure B The variance is 0.863973, 0.003274, and the 95% confidence interval is calculated as (0.857555, 0.870389) through Equations 5 and 6.

Briefly explaining the example shown in FIG. 9 again, it means that the state indicators in the case of failure A and failure B of the system correspond to (0.647169, 0.655972) and (0.857555, 0.870389), respectively.

Referring back to FIG. 8 , the failure diagnosis and prediction unit 140 uses defect information observed from the defect detection unit 130 and defects calculated in real time from the model learning unit 120 in the data reception and storage unit 110. It is possible to perform fault diagnosis and prediction by receiving condition indicators corresponding to the calculated normal state and 95% confidence intervals of condition indicators of various fault states.

10 is a graph modeled by a primary equation in the failure diagnosis and prediction unit 140 after receiving real-time sensor data from the data reception and storage unit 110 into the model learning unit 120 and calculating a state index value This is an example.

Referring to FIG. 10, the set of state indicator values calculated by receiving the sensor data received from the data reception and storage unit 110 as input to the model learning unit 120 is [0.8082714845505117, 0.7999622114525744, 0.791487347682909, 0.79958385821] 22862 , . 0.7947724620408656, 0.7896694849329334, 0.771571928580048, 0.7806675209093021, 0.7716661049792911, 0.7883671432199814, 0.76 60400044051897, 0.7767948531953544, 0.7637158382433687, 0.7827380307156276]. It can be confirmed that the set of real-time state indicator values is modeled as a linear straight line through Equation 7 below. Equation 7 below represents a general linear regression equation.

In Equation 7 above, ' 'is a straight line to estimate' ' is a parameter of ' ' is a set of state indicator values calculated by receiving the sensor data received from the data receiving and storing unit 110 into the model learning unit 120.

Referring back to FIG. 10 , it is possible to estimate the time at which the confidence interval of failure A shown in FIG.

Briefly explaining the example shown in FIG. 10 again, looking at the change trend curve of the real-time state indicator values over time, from the time when the defect starts (Time 0 in FIG. 10) to the time when the last real-time state indicator value is measured From the fact that the value decreases until (Time 20 in FIG. 10), it can be predicted that the failure A in FIG. 9 is approaching. Therefore, it can be diagnosed as a fault for failure A. In addition, the time to reach the confidence interval range of failure A from the change trend curve of the real-time condition indicator is from Time 20 to Time 100, and it is possible to diagnose that failure A will occur after Time 80 has elapsed.

11 is a graph example of a change trend of a real-time condition indicator value toward a fault B as opposed to a change trend of the real-time condition indicator toward the fault A.

Referring to FIG. 11, the real-time state indicator values are [0.7989974483369516, 0.7993925831353061, 0.804725218215658, 0.7985067208999138, 0.8011089040421506, 0.8087378 394784509, 0.8064105990638816, 0.7975970657822791, 0.8113934234882457, 0.8020159774871587, 0.8053211349719757, 0.8001695435 618309, 0.8096599029574428, 0.8029610741313947, 0.8128719008249679, 0.810044525959354, 0.8067756246385512, 0.81957648833447 55, 0.8142903098568228, 0.8173273019200041], showing an increasing trend with time from the time of occurrence of the defect. Therefore, it can be diagnosed as a defect of failure B according to the change trend of the real-time status indicator value, and the prediction time of failure occurrence can be confirmed as Time 80.

According to the change trend of the real-time condition indicator value, a change trend curve can be calculated through a high-order polynomial.

12 is an example of a graph modeled with a quadratic curve for real-time state index values.

Referring to FIG. 12, the real-time state indicator values are [0.8087958629496578, 0.8109826706967816, 0.8096827053085732, 0.8046418091755376, 0.8025164797091783, 0.799971 6881636824, 0.8010433309877496, 0.8063684514881266, 0.8021963200171204, 0.8054660190735634, 0.7913854003535287, 0.812097767 8861199, 0.7878414166072655, 0.8027761347502894, 0.8081885342052546, 0.8056805947834116, 0.7942452363793407, 0.803512191921 5729, 0.7924143768825517, 0.7928789733165137, 0.8077184167172129, 0.7855867570569314, 0.7942349852586097, 0.7887051157056041, 0.7860966541210463, 0 .7919814630631739, 0.7848281150810076, 0.7772668288531478, 0.8013310718822926, 0.7768341080553052, 0.7860801616257375, 0.78 99289003490034, 0.7850593489334408, 0.773605861462604, 0.7775930218258464, 0.7722090027980141, 0.7923964147859431, 0.769738 8305284556, 0.7887033362458561, 0.7851203997532492], depending on the time since the fault occurred. shows a rapidly declining trend. It can be seen that the real-time state index value is modeled through Equation 8 below as a quadratic curve.

In Equation 8, ' 'is the curve to estimate' It is a parameter of '.

As can be seen in FIG. 12 , as the real-time state indicator value decreases with time, it is closer to the range of the confidence interval of the failure, which can be diagnosed as a defect of failure A. The point at which the last real-time status index value is observed is Time 40, and it is predicted that failure A will be reached at Time 100, and the remaining time until failure A can be estimated at Time 60.

Using the machine learning clustering algorithm and statistical techniques described above, the method for detecting and diagnosing and predicting failures in a system without failure data is written to perform each step in a computer and recorded on a computer-printable recording medium. It can be implemented in the form of a computer program.

A computer program is a computer language such as C/C++ or Python that can be read by a processor (CPU) of the computer through a device interface of the computer so that the computer reads the program and executes the methods implemented in the program. Coded code may be included. Such codes may include functional codes related to functions defining necessary functions for executing the above methods. In addition, it may include an execution procedure control code necessary for the processor of the computer to execute the functions according to a predetermined procedure. This code may include a memory reference code indicating where additional information necessary for the processor of the computer to execute the functions should be referenced in internal or external memory of the computer. In addition, if the processor of the computer needs to communicate with any other remote computer or server in order to execute the functions, the code is how to communicate with any other remote computer or server using the communication module of the computer. It may further include communication-related codes for whether or not to do so.

The method for detecting faults and diagnosing and predicting faults in a system without fault data using the machine learning clustering algorithm and statistical techniques described herein is only one embodiment of the present invention, and various steps may be added as needed. , Since the following steps may also be performed by changing the order, the present invention is not limited to each step and its order described below. This can be equally applied to other embodiments below.

Although the embodiments of this specification have been described with reference to the accompanying drawings, those skilled in the art to which this specification belongs will understand that the present invention can be implemented in other specific forms without changing the technical spirit or essential features. You will be able to. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive.

Claims (16)

A method for detecting faults and diagnosing and predicting faults in systems without fault data using machine learning clustering algorithms and statistical techniques, a data receiving and storing unit that receives machine and facility data in real time and stores the received data;

A model learning unit that extracts features of the data received and stored in real time and applies the extracted features to a clustering algorithm of machine learning to calculate state indicators of the data;

a defect detection unit configured to set a confidence interval of the condition indicator distribution calculated for the stored steady-state data and to detect a defect in the system using the condition indicator calculated for the data received in real time;

A confidence interval of the condition indicator distribution calculated for the stored normal state data and failure state data is set, system failure is diagnosed using the condition index calculated for the real-time received data, and time from failure to failure is determined. A failure diagnosis and prediction unit that predicts

and a defect detection and failure prediction device including a display unit providing a user with defect detection and failure prediction information of the system including the machine and device.
According to claim 1,

The data receiving and storing unit,

For all systems that can acquire data through facilities and sensors including machines and devices, data that can be acquired from inside the system (system internal data) and data acquired from sensors (temperature, humidity, vibration, etc.) receiving device including sensing data).
According to claim 2,

The data receiving and storing unit,

A data reception and storage device comprising receiving real-time data from a system and storing data in normal or fault conditions.
According to claim 1,

In the fault detection method using a machine learning clustering algorithm and a statistical technique to detect faults in a system without fault data and a fault diagnosis prediction method,

Extracting a feature by receiving data from the data receiving and storing unit;

Generating normal and abnormal clusters by applying the features to a clustering algorithm of machine learning;

Calculating a cluster classification evaluation index of the classified cluster;

A fault detection and failure diagnosis and prediction method comprising the step of defining the calculated cluster classification evaluation index as a condition index.
According to claim 4,

The step of receiving data from the data receiving and storing unit and extracting features,

A method for detecting, diagnosing and predicting failures, characterized in that for extracting statistical characteristics or features related to failures from the data of the data receiving and storing unit and applying them to a clustering algorithm of machine learning.
According to claim 4,

The model learning unit,

The step of generating normal and abnormal clusters by applying the feature to a clustering algorithm of machine learning is a defect characterized in that normal and abnormal clusters are formed by applying a clustering algorithm, which is one of the unsupervised learning methods of machine learning. Detection, fault diagnosis and prediction methods.
According to claim 4,

In the step of calculating the cluster classification evaluation index of the classified cluster,

A defect detection and failure diagnosis and prediction method characterized by calculating a cluster classification evaluation index for evaluating the classification degree of the normal and abnormal clusters.

According to claim 4,

In the step of defining the calculated cluster classification evaluation index as a condition index,

The fault detection and failure diagnosis and prediction method characterized in that the calculated cluster classification evaluation index is defined as a state index representing the state of the system.
According to claim 2,

The defect detection unit,

In the fault detection method using a machine learning clustering algorithm and a statistical technique to detect faults in a system without fault data and a fault diagnosis prediction method,
Calculating a parameter of a probability distribution of state indicator values for the steady state calculated by the model learning unit;
Detecting a defect by comparing it with a condition indicator of data collected in real time using a confidence interval of the probability distribution, and
A method for detecting and diagnosing faults in a system without fault data using a machine learning clustering algorithm and a statistical technique, including providing information on whether or not there is a fault in the machine facility to a user.
According to claim 9,

Calculating the parameters of the probability distribution of the state indicator values for the steady state calculated by the model learning unit,

A defect detection and failure prediction method, characterized in that for estimating parameters by a maximum likelihood method or the like by finding a probability distribution suitable for the distribution of the condition indicators.
According to claim 9,

Determining the defect of the data collected in real time using the confidence interval of the probability distribution,

A defect detection and failure prediction method for determining that a system defect exists when the calculated state index of the data obtained in real time from the model learning unit deviates from the confidence interval of the probability distribution.
According to claim 2,

The failure diagnosis and prediction unit,

In the fault detection and fault diagnosis prediction method of a system without fault data using a machine learning clustering algorithm and statistical techniques,

Receiving defect information of the system and a status indicator at the time of the defect in the defect detection unit;

Calculating parameters of an appropriate probability distribution of state index values for the normal state and various failure states calculated by the model learning unit;

Calculating the real-time data received in the data reception and storage unit as a status indicator in the model learning unit and then defining a trend of change in the real-time status indicator;

diagnosing and predicting a failure state by comparing it with a confidence interval of each failure state according to the trend of the state indicator change of the real-time data; and

A fault detection and fault prediction method comprising providing a user with information about a type of fault to occur in the system and a predicted time.
According to claim 12,

Calculating a parameter of an appropriate probability distribution of state index values for the normal state and various failure states calculated and stored in the model learning unit finds a probability distribution suitable for the distribution of cluster classification variables for each failure state. Fault detection and failure diagnosis and prediction method characterized by estimating parameters by maximum likelihood method, etc.
According to claim 12,

Calculating the trend of the real-time condition indicator change,

Fault detection and failure diagnosis and prediction method characterized by estimating the trend of change over time of the condition indicator calculated over time using a high-order polynomial or several functions
15. The method of claim 14,

The defect detection and failure diagnosis and prediction method, characterized in that the trend of change over time of the real-time condition indicator can be estimated by a method such as regression analysis.
According to claim 12,

The step of diagnosing and predicting a failure state by comparing with the confidence interval of each failure state according to the change trend of the state indicator of the real-time data,

A fault detection and fault diagnosis and prediction method, characterized in that for diagnosing and predicting a fault from a fault by determining whether a change trend of the condition indicator is approaching a confidence interval of a certain fault state.