KR20220160992A - System for health diagnosis and remaining usefule lifetime estimation through adaptive clustering of multidimensional signals - Google Patents

Abstract

The present invention relates to a state diagnosis and life prediction system through adaptive clustering of multidimensional signals. The state diagnosis and life prediction system through adaptive clustering of multidimensional signals according to the present invention may include: a dimensional conversion unit which receives multidimensional signals and converts the signals into feature vectors to extract latent variables; a cluster prediction unit which calculates and outputs prediction strength after cluster processing for latent variables; and a state diagnosis unit which compares the prediction strength with a reference value, updates the cluster model, repeatedly performs a clustering process, and calculates a RUL diagnosis score.

Description

Condition diagnosis and life prediction system through adaptive clustering of multidimensional signals

The present invention relates to a health diagnosis and life prediction system through adaptive clustering of multidimensional signals.

As a RUL (Remaining Useful Life) prediction method according to the prior art, it is difficult to secure data for RUL analysis according to the data-based method.

In addition, according to the mathematical model-based method, there is a problem in that prediction accuracy is relatively low and it is not easy to apply the model according to the individual characteristics of facilities/devices.

As a RUL prediction method, a method of predicting the future situation based on data acquired at a previous time point and measuring the degree of facility aging using the difference between the data acquired in the future situation and the predicted data has been proposed. There is a problem in that it is difficult to automatically evaluate how large the difference between the prediction data is to determine that an abnormal situation has occurred and how much the remaining life of the system is.

The present invention is proposed to solve the above problems, and performs state diagnosis through clustering of time series data, but remaining use time of the system through cluster strength measurement ( Its purpose is to provide a system that predicts RUL).

The present invention relates to a health diagnosis and life prediction system through adaptive clustering of multidimensional signals.

A system for diagnosing conditions and predicting life through adaptive clustering of multidimensional signals according to the present invention receives multidimensional signals and converts them into feature vectors to extract latent variables, a dimensional conversion unit, and processing of clustering for latent variables to calculate prediction strength. and a cluster prediction unit that outputs the result, and a state diagnosis unit that compares the predicted strength with a reference value, updates the cluster model, repeatedly performs the clustering process, and calculates the RUL diagnosis score.

According to the present invention, there is an effect of presenting an effective index for evaluating system health even in an environment without a mathematical model or a large-scale training dataset through comparison of data cluster patterns in a healthy state and an old state.

Through state diagnosis through adaptive clustering of signals according to the present invention, system health is evaluated by continuously updating a cluster model based on the current state of the system based on the clustering prediction strength, so that the system can be automatically operated without separate external intervention. There are possible effects.

By directly outputting the RUL metric score, there is an effect of solving the above-described problem according to the prior art of converting an outlier through a difference between predicted data for system output and current output data.

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

1 illustrates a condition diagnosis and life prediction system through adaptive clustering of multidimensional signals according to an embodiment of the present invention.
2 illustrates a method for diagnosing conditions and predicting life span through adaptive clustering of multidimensional signals according to an embodiment of the present invention.
3 and 4 show examples of clustering states of multidimensional signals according to system aging according to an embodiment of the present invention.
5 illustrates an example of a signal input to a state diagnosis unit based on adaptive clustering according to an embodiment of the present invention.

The foregoing and other objects, advantages and characteristics of the present invention, and a method of achieving them will become clear with reference to the detailed embodiments described below in conjunction with the accompanying drawings.

However, the present invention is not limited to the embodiments disclosed below and may be implemented in various different forms, and only the following embodiments provide the purpose of the invention, As only provided to easily inform the configuration and effect, the scope of the present invention is defined by the description of the claims.

Meanwhile, terms used in this specification are for describing the embodiments and are not intended to limit the present invention. In this specification, singular forms also include plural forms unless specifically stated otherwise in a phrase. As used herein, “comprises” and/or “comprising” means the presence of one or more other components, steps, operations, and/or elements in which a stated component, step, operation, and/or element is present. or added.

Hereinafter, in order to help the understanding of those skilled in the art, the background in which the present invention is proposed will be described first, and then the embodiments of the present invention will be described.

System remaining time prediction technology is widely used in health management fields such as failure prediction and maintenance.

Techniques such as performing equipment abnormality diagnosis based on vibration/noise of bearings or shafts, thermal image data, etc., predicting equipment failures based on accumulated abnormalities, and determining maintenance time points are being used.

If it is possible to effectively predict the remaining operation time of system facilities, the time and cost for re-operation in the event of an accident can be greatly reduced. Therefore, interest in RUL prediction technology is increasing, but technical progress is being made due to difficulties in acquiring operation data. is a difficult situation.

That is, in order to train a model having many parameters, good quality data must be acquired, but it is very difficult to obtain data for RUL analysis.

For example, in order to acquire data for accurately predicting the remaining life of a battery, it is necessary to repeat charging and discharging of the battery tens of thousands of times and acquire the aging curve change of the charging and discharging pattern over time. There is a problem of building a sufficient number of data sets while repeating this process for each model.

As another method, a mathematical model-driven method, rather than a data-driven method, is also used, but this model-based method has relatively low prediction accuracy, and model application is difficult depending on the individual characteristics of equipment or devices. There are some difficult problems.

Therefore, in most fields where there is little life-cycle data for RUL prediction and model-based analysis is not suitable, the future situation is predicted based on the data already acquired at the present time, and the actual data and predictions acquired at the future time point The abnormality index is diagnosed by the amount of difference between data.

For example, the assumption that solar power prediction is continuously performed for facility diagnosis of solar power facilities, and the degree of facility aging can be measured as the difference between the current actual generated power amount and the predicted power amount increases over time. to be.

This method does not require much life cycle data of the system equipment and can be a good alternative if there is no suitable numerical model for RUL prediction, but the prediction accuracy for the system state must be sufficiently high, and above all, There is a problem in that it is difficult to automatically evaluate how much the difference between prediction data must be before an abnormal situation occurs and how much the remaining life of the system is.

The present invention has been proposed to solve the above problems, and provides a system for detecting a change over time between a data cluster pattern in a healthy state and a data cluster pattern in an aging state, and outputting a corresponding RUL score. .

According to the present invention, a system is provided that updates the model according to the predicted strength of the clustering pattern, increases the model accuracy, detects the accumulated system deterioration, and outputs a metric score for RUL diagnosis.

1 illustrates a condition diagnosis and life prediction system through adaptive clustering of multidimensional signals according to an embodiment of the present invention.

A system for diagnosing conditions and predicting life through adaptive clustering of multidimensional signals according to an embodiment of the present invention includes a dimension conversion unit 110 that receives multidimensional signals and converts them into feature vectors to extract latent variables, and cluster processing for latent variables. Then, a cluster prediction unit 120 that calculates and outputs the predicted strength, and a state diagnosis unit 130 that compares the predicted strength and a reference value, updates the cluster model, repeatedly performs the clustering process, and calculates the RUL diagnosis score. do.

The dimensional conversion unit 110 converts the multi-dimensional input signal into a low-dimensional feature vector to more clearly express the meaning of the data.

In general, in the case of a multidimensional input signal, since the correlation between input variables increases and causes a multi-collinearity problem, it is necessary to convert the signal into a low-dimensional feature vector and extract intrinsic feature values.

A low-dimensional feature vector containing the intrinsic characteristics of such data is called a latent variable (z).

According to an embodiment of the present invention, the dimension conversion unit 110 uses machine learning techniques such as PCA, LDA, SVD, NMF, etc., is a deep learning-based model such as an encoder of an autoencoder model, or uses several convolutional layers. It is a nested deep learning-based model.

로 계산하여 출력한다. The cluster predictor 120 calculates the clustering accuracy after cluster processing for the dimensionally reduced latent variables as the predicted strength.

Calculate and output

}, 군집별 첨도의 역수에 대한 특징값

을 출력한다. The cluster predictor 120 includes the number of clusters of latent variable data together with the predicted strength.

}, the feature value for the reciprocal of the kurtosis per cluster

outputs

The prediction strength calculation for measuring the clustering accuracy of the cluster predictor 120 is as shown in [Equation 1].

[Equation 1]

D[…] _m,n means the distance between m,n samples equal to the Euclidean distance.

로 표기한다. The case of clustering the set of latent variables X={x _p }, p=1,2,...,N at a random point in time (t) into k clusters.

marked with

은 NxN 배열을 이루며, 잠재변수 m과 n이 동일 군집에 있을 경우

=1, 동일 군집이 아닐 경우,

=0로 구성하는 co-membership 함수이다. In addition,

forms an NxN array, and if latent variables m and n are in the same cluster

= 1, if not in the same cluster,

= 0 is a co-membership function.

함수는 이전 시점(t-1)에서의 k개의 군집화된 표본과 현재 잠재 변수 간의 co-membership을 구하는 함수가 된다. in other words,

The function becomes a function that finds the co-membership between k clustered samples at the previous time point (t-1) and the current latent variable.

은 t시점에서의 k개의 군집 중 i번째 군집에서의 군집 데이터를 나타내는 인덱스이다.

is an index representing cluster data in the ith cluster among k clusters at time t.

은 t시점에서의 k개의 군집 중 i번째 군집에서의 군집 데이터 수이다.

is the number of cluster data in the ith cluster among k clusters at time t.

는 이전 시점에서의 군집 데이터와 현재 시점에서의 군집 데이터 간 동일 군집으로 할당된 잠재 변수 쌍의 수를 계산하여 제공한다. Thus, finally

calculates and provides the number of latent variable pairs assigned to the same cluster between the cluster data at the previous time point and the cluster data at the current time point.

That is, high prediction strength means that the k clustering performance models at the previous time point and the current time point were effectively performed.

이 될 수도 있으므로, 일반적으로

값을 최대로 하는 가장 큰 k를 선택하게 된다. However, when k = 1, all samples are collected in the same cluster,

can be, so in general

The largest k that maximizes the value is selected.

}, 군집별 첨도의 역수에 대한 평균, 표준편차 등의 특징값

, 그리고

값을 출력한다.The cluster predictor 120 calculates the k value at this time.

}, feature values such as mean and standard deviation for the reciprocal of kurtosis by cluster

, and

print the value

According to an embodiment of the present invention, the cluster predictor 120 uses clustering techniques such as k-means and hierarchical.

According to an embodiment of the present invention, models such as prediction strength, gap statistic, silhouette score, and elbow (scree) plot may be used as the prediction strength measurement function of the cluster prediction unit 120.

값이 기준치(th_T) 이하일 경우, 군집 모델을 재수행하여 예측강도

가 기준치를 초과하도록 군집화 프로세스를 반복적으로 수행한다. The state diagnosis unit 130

If the value is less than the reference value (th_T), the cluster model is re-performed to predict the strength

The clustering process is repeatedly performed so that A exceeds the criterion.

The state diagnosis unit 130 obtains an auxiliary weighing point for RUL through [Equation 2] and [Equation 3] below, and outputs the RUL weighing point through [Equation 4] below.

값이 기준치(th_T)이하일 경우에만 실행되는 것이 가능하다. The process repetition of the cluster model of the state diagnosis unit 130 can be automatically executed based on a certain period (T), and the prediction strength is

It is possible to execute only when the value is less than the reference value (th_T).

[Equation 2]

이다. [Equation 2] is for obtaining the RUL auxiliary metric score #1, and the ratio of the number of clusters for the latent variable at the initial time point (t ₀ ) and the number of clusters for the latent variable at the current time point (t).

to be.

값이 기준치(th_T)이하일 경우, 군집화 프로세스가 재수행되는 최초 시점-현재 시점을 의미한다. Here, the initial time point and the current time point are determined by the state diagnosis unit.

If the value is less than or equal to the reference value (th_T), it means the initial point in time at which the clustering process is re-executed - the current point in time.

As the overall system ages, the cohesion or density of the intrinsic characteristic of the multidimensional signal sensing the system decreases, and clusters with new characteristics are created.

By comparing the number of newly created clusters compared to the initial state, it is possible to evaluate system aging.

[Equation 3]

[Equation 3] is for obtaining the RUL auxiliary metric score #2, which is a metric value such as the average and standard deviation of the reciprocal of the kurtosis for each cluster at the current time point.

Kurtosis is an indicator of the sharpness of data distribution, indicating how close the distribution of observations is to the center.

In general, if the kurtosis is close to 3, it is recognized as a normal distribution, and in the case of a positive number greater than 3, it is judged as a distribution that is sharper than the normal distribution.

[Equation 3] is a mathematical equation proportional to the reciprocal of kurtosis, and the equation was converted to indicate a larger number as the kurtosis is lower.

As the overall system ages, the kurtosis in the cluster distribution decreases as the cohesion or density of the intrinsic characteristic of the multidimensional signal sensing the system decreases.

It is possible to evaluate system oldness by comparing the kurtosis of the clusters in the initial state and the kurtosis of newly created clusters.

For example, the quantification of the characteristic value for the reciprocal of the kurtosis for each cluster at the current time point may use the average and standard deviation in [Equation 3] or use various indicators such as the minimum value, maximum value, and sum of cluster kurtosis. .

를 통해 계산될 수 있다.For example, there may be a system in which the kurtosis of each cluster at the current time point is greater than at the previous time point, and in this system, RUL ₂ is

can be calculated through

[Equation 4]

[Equation 4] represents the RUL metric score (Y).

[Equation 2] and [Equation 3] are ensemble to derive the final metric score.

At this time, the value of α is a parameter having a value greater than 0 and less than 1.

For example, RUL ₁ and RUL ₂ may be normalized values.

For example, the value of (1-α) may be replaced by another variable β.

2 illustrates a method for diagnosing conditions and predicting life span through adaptive clustering of multidimensional signals according to an embodiment of the present invention.

Step S210 receives a multi-dimensional signal.

Step S220 converts the multidimensional signal input value (X) of the system into a value of a latent variable (z) including intrinsic characteristics through a dimension converter.

In step S230, the number of optimally clustered data clusters, cluster prediction strength, and reciprocal of kurtosis (data distribution shape) for each cluster are obtained through latent variable-based data cluster model learning.

At this time, the cluster strength prediction technique such as [Equation 1] can be used for the optimal clustering value.

As a result of the comparison in step S240, if the output of the optimal cluster model is smaller than the predetermined estimated value of the cluster strength, the cluster model is updated to improve the performance of predicting the cluster strength.

At this time, the new cluster generation rate (RUL ₁ ) and the change in the data distribution shape within the cluster (RUL ₂ ) occur due to the renewal of the cluster model. In step S250, the RUL metric score (Y) is obtained based on this.

3 and 4 show examples of clustering states of multidimensional signals according to system aging according to an embodiment of the present invention.

Referring to FIG. 3 , clusters having new characteristics are created as the inherent characteristic value distribution cohesion or density of the multidimensional signal sensing system changes as the system ages, and the kurtosis in the new cluster distribution also changes.

Referring to FIG. 3, as the data distribution with high kurtosis in three clusters at time t ₀ passes through time points t ₁ and t ₂ , new clusters are created due to changes in the intrinsic characteristic data distribution, and the kurtosis for each cluster is also It changes.

Referring to FIG. 4 , as the system ages, the coherence or density of the inherent characteristic value distribution of the multidimensional signal sensing system changes, and outliers frequently occur, and clusters having new characteristics are sequentially generated.

Referring to FIG. 4, 3 clusters at time t ₀ are changed to 4 and 5 clusters due to changes in the distribution of intrinsic characteristic data as time passes t ₁ , t _2, etc., and due to aging over time The data cluster gradually changes compared to the initial cluster.

5 illustrates an example of a signal input to a state diagnosis unit based on adaptive clustering according to an embodiment of the present invention.

군집예측강도 저하되고, 이에 따라 군집모델이 갱신되면서

와

신호가 함께 변화하는 상태 진단부(130)에 대한 입력 신호의 예를 도시한다. Referring to FIG. 5, according to the change in the intrinsic characteristic value distribution cohesion or density of the multidimensional signal sensing the system according to system aging, while sequentially passing through time points t _{0 ,} t ₁ , and t ₂ , due to changes in the distribution of intrinsic characteristic data

As the strength of cluster prediction decreases, the cluster model is updated accordingly.

Wow

An example of an input signal to the state diagnosis unit 130 in which the signal changes together is shown.

Meanwhile, the method for diagnosing conditions and predicting lifespan through adaptive clustering of multidimensional signals according to an embodiment of the present invention may be implemented in a computer system or recorded on a recording medium. A computer system may include at least one processor, a memory, a user input device, a data communication bus, a user output device, and a storage. Each of the aforementioned components communicates data through a data communication bus.

The computer system may further include a network interface coupled to the network. The processor may be a central processing unit (CPU) or a semiconductor device that processes instructions stored in memory and/or storage.

The memory and storage may include various types of volatile or non-volatile storage media. For example, memory may include ROM and RAM.

Therefore, the method for diagnosing conditions and predicting lifespan through adaptive clustering of multidimensional signals according to an embodiment of the present invention can be implemented as a method executable on a computer. When the method for diagnosing conditions and predicting lifespan through adaptive clustering of multidimensional signals according to an embodiment of the present invention is executed in a computer device, computer readable instructions are provided for diagnosing conditions and predicting life through adaptive clustering of multidimensional signals according to an embodiment of the present invention. Life expectancy methods can be performed.

Meanwhile, the above-described method for diagnosing conditions and predicting life through adaptive clustering of multidimensional signals according to the present invention can be implemented as computer readable codes on a computer readable recording medium. Computer-readable recording media includes all types of recording media in which data that can be decoded by a computer system is stored. For example, there may be read only memory (ROM), random access memory (RAM), a magnetic tape, a magnetic disk, a flash memory, an optical data storage device, and the like. In addition, the computer-readable recording medium may be distributed in computer systems connected through a computer communication network, and stored and executed as readable codes in a distributed manner.

Claims (1)

a dimensional conversion unit that receives multidimensional signals and converts them into feature vectors to extract latent variables;
a cluster prediction unit that calculates and outputs prediction strength after cluster processing for the latent variables; and
Condition diagnosis unit that compares the predicted strength and reference value, updates the cluster model, repeatedly performs the clustering process, and calculates the RUL diagnosis score
Condition diagnosis and life prediction system through adaptive clustering of multi-dimensional signals including.

Images