KR20230075826A - Deep-learning based fault detection apparatus and method for motor - Google Patents

Abstract

The present invention discloses a deep learning-based motor failure detection device and method. Conventional motor diagnosis technologies used a simple method of using a single sensor to determine a failure when the average, maximum, or minimum value of the sensor exceeds a threshold, but the present invention utilizes various sensor data and utilizes wavelet features that can reflect not only the statistical characteristics of the data but also the time series characteristics, thereby enabling more precise diagnosis. In addition, the present invention can process various sensor data at once, and because the present invention uses a deep learning-based failure diagnosis method, accurate failure diagnosis is possible through model learning even if the data has non-linear characteristics. Due to the detailed effects described above, the present invention can perform motor failure diagnosis in real time for all systems and equipment using motors, so unplanned stoppage of systems and equipment can be prevented, thereby increasing productivity and economic effectiveness. In addition, because many industrial sites carry out various tasks, considering the reality that systems and equipment are not operated under a single condition, the deep learning-based failure detection method of the present invention is advantageous for multimodal data processing. Therefore, the method is particularly suitable for failure diagnosis of spindle motors whose operating characteristics vary depending on the operating environment.

Description

Deep-learning based fault detection apparatus and method {Deep-learning based fault detection apparatus and method for motor}

The present invention relates to a method and system for detecting a motor failure, and more particularly, to a device and method for detecting a motor failure based on deep learning.

Various methods for diagnosing failures of motors used in machine tools have been proposed, and recently, methods for diagnosing failures using artificial intelligence have been proposed.

In the prior art, features are extracted using current data or vibration data individually, and spindle motors are diagnosed using this, and most models use simple frequency conversion or supervised learning-based I am only using the model.

However, if only a single data is used, it is impossible to cope with various failures, and diagnosis is possible only for specific failures or abnormalities, so various sensors must be used.

In addition, in general industrial processes, it is very difficult to obtain failure data because maintenance is performed before a system or equipment fails. Therefore, supervised learning-based classification models that require a large number of failure data are not suitable for industrial process failure diagnosis.

In addition, general frequency conversion methods or methods using statistical characteristics (Kurtosis, Skewness, etc.) may not be suitable for diagnosing motors that perform continuous machining because they cannot reflect time-series data characteristics, so an algorithm that can reflect time-series characteristics is needed. The situation is.

An object of the present invention is to provide a deep learning-based motor failure detection device and method capable of performing failure diagnosis using steady state data collected while the motor is operating in a steady state.

A motor failure detection method according to a preferred embodiment of the present invention for solving the above problems is a motor failure detection method performed in a motor failure detection apparatus, (a) deep using feature vectors collected from a motor in a normal state. learning a learning-based learning model and setting a decision threshold; (b) receiving feature vectors from a motor to be detected; (c) generating a predictive value by inputting the input feature vectors into a diagnosis model to which the structure of the learning model and hyperparameters are applied; (d) generating a diagnostic index using an error between the predicted value and the input feature vectors; and (e) determining whether the motor is out of order by comparing the diagnostic index with the determination threshold.

In addition, the step (d) may include (d1) calculating an error between corresponding items of the collected feature vectors and the predicted value; and (d2) generating a diagnostic index by summing the squares of errors for each item.

Also, in the step (e), it may be determined that a failure has occurred if the diagnostic index is greater than the determination threshold value for more than a predefined time period.

In the step (c), re-learning is performed on the diagnostic model to which the learning model structure and hyperparameters are applied, and collected feature vectors may be input to the diagnostic model.

In addition, in step (a), learning and verification can be performed by sequentially performing k -fold cross-validation and leave-one-out verification on the learning model using feature vectors collected from motors in a steady state. .

In addition, in step (a), learning and verification are performed by sequentially performing k -fold cross-validation and leave-one-out verification on the learning model using the feature vectors collected from the motor in (a1) the steady state. doing; (a2) inputting feature vectors collected from motors in a steady state to the learning model, calculating a prediction error between a predicted value output from the learning model and the input feature vectors, and calculating a squared prediction error; and (a3) using a normal distribution of squared prediction errors obtained by performing steps (a1) and (a2) on a plurality of feature vectors, using a confidence limit of a predefined critical confidence interval as the decision threshold It may include setting steps.

Meanwhile, a computer program according to a preferred embodiment of the present invention for achieving the above object is stored in a non-temporary storage medium and executed in a computer including a processor to perform the motor failure detection method.

On the other hand, a motor failure detection device according to a preferred embodiment of the present invention for achieving the above object is a motor failure detection device including a processor and a memory for storing predetermined instructions, the processor executing the instructions stored in the memory (a) learning a deep learning-based learning model using feature vectors collected from motors in a steady state, and setting a decision threshold; (b) receiving feature vectors from a motor to be detected; (c) generating a predictive value by inputting the input feature vectors into a diagnosis model to which the structure of the learning model and hyperparameters are applied; (d) generating a diagnostic index using an error between the predicted value and the input feature vectors; and (e) determining whether the motor is out of order by comparing the diagnostic index with the determination threshold.

In addition, the step (d) may include (d1) calculating an error between corresponding items of the collected feature vectors and the predicted value; and (d2) generating a diagnostic index by summing the squares of errors for each item.

Also, in the step (e), the processor may determine that a failure has occurred if a time period in which the diagnosis index is greater than the determination threshold value occurs for a predetermined period of time or longer.

In the step (c), the processor may perform re-learning on the diagnostic model to which the structure of the learning model and hyperparameters are applied, and then input the collected feature vectors to the diagnostic model.

In addition, in the step (a), the processor sequentially performs k -fold cross-validation and leave-one-out verification on the learning model using feature vectors collected from motors in a steady state to perform learning and verification. can be done

In addition, in step (a), learning and verification are performed by sequentially performing k -fold cross-validation and leave-one-out verification on the learning model using the feature vectors collected from the motor in (a1) the steady state. doing; (a2) inputting feature vectors collected from motors in a steady state to the learning model, calculating a prediction error between a predicted value output from the learning model and the input feature vectors, and calculating a squared prediction error; and (a3) using a normal distribution of squared prediction errors obtained by performing steps (a1) and (a2) on a plurality of feature vectors, using a confidence limit of a predefined critical confidence interval as the decision threshold It may include setting steps.

Conventional motor diagnosis technologies use a simple method of determining a failure when the average, maximum, or minimum value of a sensor exceeds a threshold using a single sensor, but the present invention utilizes various sensor data and uses not only statistical characteristics of the data, but also More precise diagnosis is possible by utilizing wavelet features that can also reflect time series characteristics.

In addition, since the present invention can process various sensor data at once and uses a deep learning-based fault diagnosis method, accurate fault diagnosis is possible through model learning even if the data has nonlinear characteristics.

Due to the detailed effects described above, the present invention can perform motor failure diagnosis of all systems and equipment using motors in real time, thereby preventing unplanned stoppage of systems and equipment, thereby increasing productivity and economic effects. can make it In addition, considering the reality that many industrial sites do not operate systems and equipment under a single condition because they perform various tasks, the deep learning-based fault detection method of the present invention is advantageous for multimodal data processing, and thus the operating environment. It is particularly suitable for fault diagnosis of spindle motors whose operating characteristics vary according to

1 is a diagram showing the overall configuration of a deep learning-based motor failure detection device according to a preferred embodiment of the present invention.
2 is a diagram illustrating an example in which a sensor is installed in a spindle motor according to a preferred embodiment of the present invention.
3 is a functional block diagram illustrating functions performed by a processor when a program stored in a memory is executed in the processor according to a preferred embodiment of the present invention.
4 is a flowchart illustrating a method for detecting a motor failure based on deep learning according to a preferred embodiment of the present invention.
5 is a diagram illustrating a process of extracting a feature vector from a sensor measurement value of a motor according to a preferred embodiment of the present invention.
6 is a diagram illustrating a k-fold cross validation method.
7 is a diagram illustrating a leave-one-out verification method.
8 is a diagram illustrating a process of determining whether a motor is out of order in a preferred embodiment of the present invention.

Hereinafter, preferred embodiments of the present invention will be described with reference to the drawings.

Here, the above objects, features and advantages of the present invention will become more apparent through the following detailed description in conjunction with the accompanying drawings. However, the present invention can apply various changes and can have various embodiments, hereinafter, specific embodiments will be illustrated in the drawings and described in detail.

In the drawings, the thickness of layers and regions is exaggerated for clarity, and elements or layers may be "on" or "on" other elements or layers. What is referred to includes all cases where another layer or other component is intervened in the middle as well as immediately above another component or layer. Like reference numerals designate essentially like elements throughout the specification. In addition, components having the same function within the scope of the same idea appearing in the drawings of each embodiment are described using the same reference numerals.

If it is determined that a detailed description of a known function or configuration related to the present invention may unnecessarily obscure the subject matter of the present invention, the detailed description will be omitted. In addition, numbers (eg, 1st, 2nd, etc.) used in the description process of this specification are only identifiers for distinguishing one component from another component.

1 is a diagram showing the overall configuration of a deep learning-based motor failure detection device according to a preferred embodiment of the present invention, and FIG. 2 is a diagram illustrating an example in which a sensor is installed in a spindle motor according to a preferred embodiment of the present invention. am.

1 and 2, the deep learning-based motor failure detection device 100 according to a preferred embodiment of the present invention includes a vibration sensor 121, a current sensor 123, a temperature sensor 125, and a processor 110. ), a memory 130, an input means 150 and an output means 170.

The memory 130 according to a preferred embodiment of the present invention may store instructions executable by the processor 110 and programs executed by the processor 110, and may store input and output data.

The processor 110 according to a preferred embodiment of the present invention executes instructions stored in the memory 130 to perform each step of a deep learning-based motor failure detection method described later with reference to FIG. 4 . The memory 130 may be alternatively operated as a web storage or cloud server that performs the function of a storage medium on the Internet.

The input means 150 is implemented as a typical input device such as a mouse and keyboard, and can receive setting information and selection information from a user and output them to the processor 110 .

The output unit 170 may be implemented as a monitor or the like to display data generated by the processor 110 to a user.

Vibration sensors 121 are installed at a plurality of locations of the motor to detect vibration generated by the motor and output the result to the processor 110, and temperature sensors 125 are also installed at a plurality of locations of the motor to detect vibration generated by the motor. The temperature is measured and output to the processor 110 . The current sensor 123 measures the current flowing in the stator of the motor and outputs it to the processor 110 .

Although the present invention can be applied to a general motor, for convenience of explanation, a case of detecting a failure of a spindle motor will be described as an example.

3 is a functional block diagram representing functions performed by a processor when a program stored in a memory is executed in a processor according to a preferred embodiment of the present invention, and FIG. 4 is a deep learning-based motor according to a preferred embodiment of the present invention. It is a flow chart explaining the failure detection method.

Referring to FIGS. 3 and 4 , when a program stored in a memory is loaded into a processor and executed, the processor conceptually includes an offline learning module and an online diagnosis module. The offline learning module includes a deep learning learning model 111, a first error calculator 112-1, a first SPE calculator 113-1, and a threshold setting unit 114. The online diagnosis module includes a diagnosis model 115, a second error calculator 112-2, a second SPE calculator 113-2, and a decision unit 116.

First, the off-line learning module obtains biases and weights to be applied to the diagnostic model 115 included in the diagnostic module using past data collected when the motor operates in a steady state, and sets a threshold value.

The online learning module detects a failure by applying data collected in real time from a motor that is a failure diagnosis target to the diagnosis model 115 .

Hereinafter, the function of each detailed component will be described sequentially. Feature vectors extracted from data collected from a motor operating in a steady state are input to the deep learning learning model 111 in the form of a training data matrix.

First, in a preferred embodiment of the present invention, a feature vector is input to the deep learning learning model 111 and the diagnosis model 115. The feature vector is a feature extracted from data measured from each sensor, and intermediate to obtain statistical characteristics. Value, average value, standard deviation, etc. are calculated, and wavelet transformation is applied to extract frequency characteristics of time series data to obtain a feature vector (X) as shown in FIG. 115).

The deep learning learning model 111 is implemented as a semi-supervisored learning model such as an auto-encoder and AAKR (Auto Associative Kernel Regression). The auto-encoder and AAKR learning model operate in a steady state. Learning proceeds using the data collected during the process.

In the case of an autoencoder, the input data is re-constructed by reducing and expanding the dimension of the input measured value vector, and in the case of AAKR, the input data is obtained by calculating a weighted average using the distance to the input measured value. rebuild

In these two algorithms, if the input measurement values have a similar pattern to the steady-state driving data, values similar to the input measurement values will be output from the pattern learning model, and patterns with low similarity will result in a lot of error with the input measurement values. Since the two deep learning models themselves are known algorithms, detailed descriptions are omitted.

On the other hand, the deep learning learning model 111 receiving the training data matrix is trained and verified in a k -fold cross validation method, and then learns and validates in a leave-one-out method. Verification is performed (S411).

The k -fold cross validation method is one of the most used methods for selecting the structure and hyperparameters of the learning model 111, and as shown in FIG. 6, the entire learning dataset D is k It is divided into k data sets, and learning is performed by using one of the k data sets as verification data and using the remaining k -1 data sets as learning data. Next, the verification dataset is replaced with another one, and learning is performed using the remaining k -1 datasets as training data. In this way, when all datasets are used as a dataset for verification by performing k times, k verification results are derived. When verification is completed up to the k -th dataset, the structure or hyperparameter of the model is changed and the test is performed again, and the structure and hyperparameter with the lowest error are selected by calculating the average error.

The leave-one-out verification method is a method for selecting the threshold value of the diagnostic model 115. As shown in FIG. 7, one data d _n from the entire dataset D is used as verification data and the remaining data is used as training data. It is an iterative way of learning and testing.

이 입력되면, 이에 대응되는 예측 데이터

을 출력한다. 본 발명의 실시예에서, 학습 데이터

은 벡터 행렬 형식으로 입력되고, 예측 데이터

역시 입력 데이터에 각각 대응되는 벡터 행렬 형식으로 출력된다.The deep learning learning model 111 learned through the above process is the learning data

If is input, the corresponding predicted data

outputs In an embodiment of the present invention, learning data

is input in the form of a vector matrix, and the predicted data

It is also output in the form of a vector matrix corresponding to each input data.

를 입력받고, 딥러닝 학습 모델(111)로 입력된 학습 데이터

을 각각 입력받고, 입력 데이터와 예측치 간의 예측 오차

를 계산하여 제 1 SPE 계산부(113-1)로 출력한다.The first error calculation unit 112-1 is a predicted value from the deep learning learning model 111

, and training data input to the deep learning learning model (111)

receive each input, and the prediction error between the input data and the predicted value

Calculate and output to the first SPE calculator 113-1.

를 입력받고, 예측 오차 제곱의 합(SPE:Square Prediction Error)을 계산하여 임계치 설정부(114)로 출력한다(S412). 여기서, 예측 오차

는 행렬 형식으로 입력될 수 있고, 이 경우, 예측 오차(E)는 아래의 수학식 1과 같이 표현될 수 있으며, 진단 모델(115)의 진단지수로 활용된다.The first SPE calculator 113-1 predicts error

is received, and a square prediction error (SPE) is calculated and outputted to the threshold setting unit 114 (S412). Here, the prediction error

may be input in a matrix form, and in this case, the prediction error (E) may be expressed as in Equation 1 below, and is used as a diagnostic index of the diagnostic model 115.

The threshold value setting unit 114 sets the decision threshold SPEα of the diagnostic model 115 using the prediction error E and outputs it to the determination unit 116 (S413). The threshold setting unit 114 collects SPE results calculated when steady-state data is input to the learned model, and sets a threshold using the distribution of the data. For example, the 99th percentile of the SPE data may be used, or a confidence limit of a 99% confidence interval of a normal distribution may be set as the decision threshold.

On the other hand, the diagnosis model 115 of the online diagnosis module is a model structure from the deep learning learning model 111, hyperparameters such as bandwidth parameters, and bias (b) and weights (for transfer learning) w), etc., and re-learning is performed.

)를 생성하여 제 2 오차 계산부(112-2)로 출력한다(S430).Thereafter, the diagnostic model 115 receives a query vector (X(k)) composed of feature vectors extracted from real-time measurement values of the motor to be detected in real time (S420), and predicts values (S420).

) is generated and output to the second error calculator 112-2 (S430).

)를 질의 벡터(X(k))와 비교하여 오차 벡터(

)를 계산하여 제 2 SPE 계산부(113-2)로 출력한다.The second error calculator 112-2 calculates the predicted value (

) with the query vector (X(k)), the error vector (

) is calculated and output to the second SPE calculator 113-2.

The second SPE calculation unit 113-2 calculates the sum of square prediction errors (SPE:Square Prediction Error) (SPE(k)) used as the diagnosis index of the motor diagnosis model 115 using the error vector, It is output to the determination unit 116 (S440). Here, SPE(k) can be calculated using Equation 1 above.

After that, the determination unit 116 compares the input diagnosis index SPE(k) with the determination threshold value SPEα input from the threshold value setting unit 114 to determine whether the motor is out of order (S450).

Referring to FIG. 8 , the determination unit 116 compares the diagnostic index SPE(k) input from the second SPE calculator 113-2 with a determination threshold value indicated in red, and determines the diagnostic index SPE(k). If a time greater than the threshold occurs for more than a predefined time, it is determined that a failure has occurred.

The method for detecting a motor failure based on deep learning according to a preferred embodiment of the present invention described so far may be implemented as a computer program executable on a computer and stored in a non-temporary storage medium.

The storage medium includes all types of recording devices in which data that can be read by a computer system is stored. Examples of computer-readable storage media include ROM, RAM, CD-ROM, magnetic tape, floppy disk, and optical data storage devices. In addition, the computer-readable storage medium is distributed in computer systems connected through a network, so that computer-readable codes can be stored and executed in a distributed manner.

So far, the present invention has been looked at with respect to its preferred embodiments. Those skilled in the art to which the present invention pertains will be able to understand that the present invention can be implemented in a modified form without departing from the essential characteristics of the present invention. Therefore, the disclosed embodiments should be considered from an illustrative rather than a limiting point of view. The scope of the present invention is shown in the claims rather than the foregoing description, and all differences within the equivalent scope will be construed as being included in the present invention.

100: motor failure detection device
121: vibration sensor 123: current sensor
125: temperature sensor
110: processor 111: deep learning learning model
112-1: first error calculation unit 112-2: second error calculation unit
113-1: 1st SPE calculation unit 113-2: 2nd SPE calculation unit
114: threshold setting unit 115: diagnostic model
116: determination unit 130: memory
150: input means 170: output means

Claims (13)

A motor failure detection method performed in a motor failure detection device,
(a) learning a deep learning-based learning model using feature vectors collected from motors in a steady state, and setting a decision threshold;
(b) receiving feature vectors from a motor to be detected;
(c) generating a predictive value by inputting the input feature vectors into a diagnosis model to which the structure of the learning model and hyperparameters are applied;
(d) generating a diagnostic index using an error between the predicted value and the input feature vectors; and
(e) comparing the diagnostic index with the determination threshold to determine whether or not the motor has failed. The method of claim 1, wherein step (d) is
(d1) calculating an error between the collected feature vectors and corresponding items of the predicted value; and
(d2) generating a diagnostic index by summing the squares of errors for each item. The method of claim 1, wherein step (e)
A method for detecting a motor failure, characterized in that it is determined that a failure has occurred when the diagnostic index is larger than the determination threshold value for more than a predefined time period. The method of claim 1, wherein step (c) is
After performing re-learning on the diagnostic model to which the structure and hyperparameters of the learning model are applied, the collected feature vectors are input to the diagnostic model. The method of claim 1, wherein step (a) is
A motor failure detection method characterized in that learning and verification are performed by sequentially performing k -fold cross-validation and leave-one-out verification on the learning model using feature vectors collected from the motor in a steady state. The method of claim 5, wherein step (a)
(a1) performing learning and verification by sequentially performing k -fold cross-validation and leave-one-out verification on the learning model using feature vectors collected from motors in a steady state;
(a2) inputting feature vectors collected from motors in a steady state to the learning model, calculating a prediction error between a predicted value output from the learning model and the input feature vectors, and calculating a squared prediction error; and
(a3) Using a normal distribution of squared prediction errors obtained by performing steps (a1) and (a2) on a plurality of feature vectors, setting the confidence limit of a predefined critical confidence interval as the decision threshold Motor failure detection method comprising the step of doing. A computer program stored in a non-transitory storage medium and executed in a computer including a processor to perform the motor failure detection method according to any one of claims 1 to 6.
A motor failure detection device including a processor and a memory for storing predetermined instructions,
The processor executing the instructions stored in the memory
(a) learning a deep learning-based learning model using feature vectors collected from motors in a steady state, and setting a decision threshold;
(b) receiving feature vectors from a motor to be detected;
(c) generating a predictive value by inputting the input feature vectors into a diagnosis model to which the structure of the learning model and hyperparameters are applied;
(d) generating a diagnostic index using an error between the predicted value and the input feature vectors; and
(e) comparing the diagnostic index with the determination threshold to determine whether the motor has failed. The method of claim 8, wherein step (d)
(d1) calculating an error between the collected feature vectors and corresponding items of the predicted value; and
(d2) generating a diagnostic index by summing the squares of errors for each item. The method of claim 8, wherein in step (e)
Wherein the processor determines that a failure has occurred when a time period in which the diagnostic index is greater than the determination threshold value occurs for more than a predefined time period. The method of claim 8, wherein in step (c)
The processor performs re-learning on the diagnostic model to which the structure and hyperparameters of the learning model are applied, and then inputs the collected feature vectors to the diagnostic model. The method of claim 1, wherein in step (a)
The processor performs learning and verification by sequentially performing k -fold cross-validation and leave-one-out verification on the learning model using feature vectors collected from the motor in a steady state. Device. 13. The method of claim 12, wherein step (a)
(a1) performing learning and verification by sequentially performing k -fold cross-validation and leave-one-out verification on the learning model using feature vectors collected from motors in a steady state;
(a2) inputting feature vectors collected from motors in a steady state to the learning model, calculating a prediction error between a predicted value output from the learning model and the input feature vectors, and calculating a squared prediction error; and
(a3) Using a normal distribution of squared prediction errors obtained by performing steps (a1) and (a2) on a plurality of feature vectors, setting the confidence limit of a predefined critical confidence interval as the decision threshold Motor failure detection device characterized in that it comprises the step of doing.

Images