KR102027389B1 - Fault diagnosis system of mechanical devices using autoencoder and deep-learning - Google Patents

Abstract

The present invention relates to a machine equipment failure diagnosis apparatus using an auto-encoder and deep learning, which comprises: an input unit obtaining and storing multiple types of input data generated from a plurality of parts of mechanical equipment; and an artificial neural network including an auto-encoder converting the multiple types of input data transferred from the input unit into a feature map, a deep learning model unit previously trained with the feature map and failure diagnosis information, and an output unit inputting the feature map in the deep learning model unit and outputting a confidence degree of failures per category for the plurality of parts of the mechanical equipment.

Description

Fault diagnosis system of mechanical devices using autoencoder and deep-learning}

The present invention relates to an apparatus for diagnosing mechanical equipment failure using an autoencoder and deep learning.

 Artificial intelligence learns data from computers so that they can make their own decisions as if they were humans.

For example, to create a classification model that allows you to look at a picture and identify which picture, you need to keep showing the picture to the computer and learn what the picture is about.

This method is called 'supervised machine learning'. Deep learning, which has recently gained a lot of attention and performance, is similar to machine learning.

Deep learning and machine learning have in common that they use data to train models, but there are big differences in the process of learning data.

In order to recognize an image by machine learning, it is necessary to find a characteristic factor that can best distinguish the objects in the picture, rather than using the picture as it is.

What determines the performance of a model in machine learning is how well this characteristic factor represents the data.

Deep learning, on the other hand, does not select a feature factor, but rather finds a feature factor that can achieve the goal well in the process of learning the model from the data. In the mechanical field, finding characteristic factors from data obtained from complex mechanical systems requires a lot of expert knowledge, which makes it difficult to apply machine learning. Deep learning, on the other hand, detects feature factors automatically, and there are increasing attempts to use deep learning in the field of machine failure diagnosis.

On the other hand, mechanical and electronic equipment is exposed to the risk of failure, and the longer the use time, the higher the risk of failure.

Failure of mechanical equipment in a manufacturing plant or processing plant will greatly affect the productivity and profitability of the product, lead to industrial accidents, high costs for follow-up measures, and inconvenient use when needed. . Surprisingly, however, the cause of the failure of the mechanical equipment is difficult to identify, and even though the failure cause (part defect) configured in the mechanical equipment remains, the machine equipment is simply replaced in the hope that it will last long.

In other words, in order to diagnose the failure of the mechanical equipment, it is necessary to identify the cause of the failure (part defect) of the components configured in the mechanical equipment. There was a problem that it was difficult to pinpoint the problem, so in the industrial field, there was a problem of identifying the cause of the failure of the mechanical equipment based on the expert's empirical analysis.

In order to solve this problem, Korean Patent Laid-Open Publication No. 1996-0018877 connects them between a known controller that processes various controls as an electrical signal and an electric device / hydraulic device, and converts various electric signals into digital signals. An intermediate module comprising a unit, a communication unit for communicating with the main body, a memory for storing the various controls and signals, a CPU for performing various processes, a screen for displaying results of various signals processed by the intermediate module, and an operator. It consists of a main body including a keyboard part for the fault diagnosis apparatus to recognize the intention of the device, outputs the modem selection menu screen, outputs the function item selection menu screen according to the instruction, and measures the function according to the selected function item. Output the result to the main body screen by judging, and determine the stop guide request, help request, and save the result contents in order. Trouble-shooting systems for machinery have been published that feature needle stitching, help indications, and memory stores to process and return to the beginning.

Korean Patent Publication No. 1996-0018877 (1996.06.17)

The present invention provides a technique for solving the above problems, and an object of the present invention is to use an auto-encoder and an artificial neural network in which a deep learning model is built, and an auto-encoder capable of accurately diagnosing the type and cause of failure of a component of a mechanical equipment It is to provide a device diagnosis device using deep learning.

Mechanical equipment failure diagnosis apparatus using the auto-encoder and deep learning according to the present invention includes an input unit 100 for obtaining and storing a plurality of input data generated from a plurality of components of the mechanical equipment; And an auto-encoder 210 for converting a plurality of input data transferred from the input unit 100 into a feature map, a deep learning model unit 220 previously trained with a feature map and failure diagnosis information, and the deep learning. And a neural network 200 including an output unit 230 for inputting the feature map to the model unit 220 and outputting a degree of failure per category for a plurality of parts of mechanical equipment.

In addition, the auto encoder 210 includes an input layer unit 211 in which input data transferred from the input unit 100 is converted into an input layer, and a hidden layer in which the input layer is encoded and compressed into a hidden layer having a feature map. A unit 212, and an output layer unit 213 that the hidden layer is decoded and reconstructed as an output layer.

In addition, the autoencoder 210 calculates and controls the average squared error between the input layer and the output layer to be minimized.

The input data also includes vibration frequencies and vibration spectra that occur in multiple components of the machine.

In addition, the band pass filter 300 for gradually filtering the vibration spectrum specific section of the input data until a defect signal according to the failure diagnosis information appears in the input data; And an envelope 400 for modeling a change pattern of vibration spectrum of the input data transferred from the band pass filter 300.

In addition, the artificial neural network 200 outputs a probability value calculated by the output unit 230 using the softmax regression function as the output confidence level, and sets a target value of the output value between the output value and the target value. It includes a correction unit 240 for correcting the error by gradient descent (Gradient Descent).

In addition, the input data is data generated from a plurality of sensors separately installed in a plurality of parts of the mechanical equipment.

The input data may include data of an acceleration sensor, data of a noise sensor, data of a light sensor, data of a gyro sensor, data of a humidity sensor, data of a magnetic sensor, data of an input sensor, data of a pressure sensor, and data of a temperature sensor. At least one of the.

Accordingly, the mechanical equipment failure diagnosis apparatus using the auto encoder and the deep learning according to the present invention is an advantage that can accurately diagnose the failure type and cause of the components of the mechanical equipment by using an artificial neural network with the auto encoder and the deep learning model There is this.

1 is a block diagram showing an apparatus for diagnosing failure of a mechanical equipment using an auto encoder and deep learning according to the present invention.
Figure 2 is a schematic diagram showing a failure diagnosis process of the mechanical equipment of the mechanical equipment failure diagnosis apparatus using the auto encoder and deep learning in accordance with the present invention.
3 is a schematic diagram showing an encoding process of an autoencoder according to the present invention.
Figure 4 is a block diagram showing an example of an apparatus for diagnosing failure of mechanical equipment using the auto encoder and deep learning in accordance with the present invention.
FIG. 5 is a graph listing three examples of an envelope of an apparatus for diagnosing a failure of a mechanical equipment using an auto encoder and deep learning according to the present invention; FIG.
Figure 6 is a schematic diagram showing the failure diagnosis process of the acceleration machine equipment in the vibration frequency and vibration spectrum input to the input data of the mechanical equipment failure diagnosis apparatus using the auto-encoder and deep learning in accordance with the present invention.
Figure 7 is a schematic diagram showing a failure diagnosis process of the acceleration machine equipment in the state that the data generated from the multiple sensors to the input data of the mechanical equipment failure diagnosis apparatus using the auto encoder and deep learning according to the present invention.

Hereinafter, the technical spirit of the present invention will be described in more detail with reference to the accompanying drawings.

The accompanying drawings are only examples to illustrate the technical idea of the present invention in more detail, and thus the technical idea of the present invention is not limited to the forms of the accompanying drawings.

1 is a block diagram showing a mechanical equipment failure diagnosis apparatus using an auto encoder and deep learning according to the present invention, Figure 2 is a failure diagnosis of the mechanical equipment of the mechanical equipment failure diagnosis apparatus using an auto encoder and deep learning according to the present invention Schematic diagram showing the process.

As shown in FIGS. 1 and 2, the apparatus for diagnosing mechanical equipment failure using the auto encoder and the deep learning according to the present invention includes an input unit 100 and an artificial neural network 200.

The input unit 100 is an interface, and a plurality of input data generated from a plurality of components of the mechanical equipment can be input. At this time, the input data is stored in time series.

The artificial neural network 200 outputs a degree of failure per category for a plurality of parts of the machine by learning and analyzing the failure diagnosis information of the plurality of parts of the machine and the feature map from which the features of the input data are extracted in advance. It includes an auto encoder 210, a deep learning model unit 220, and an output unit 230.

The auto encoder 210 converts a plurality of input data transferred from the input unit 100 into a feature map. In this case, the feature map is a feature of the input data, and is a matrix consisting of a compound product calculated by filtering the input data. The feature map has fewer nodes than the input data, but is restored to its original value. can do.

The deep learning model unit 220 is trained in advance with a feature map and fault diagnosis information, and finds that the fault confidence level for each of the parts of the machine equipment is well achieved in the feature map.

The output unit 230 inputs the feature map to the deep learning model unit 220 and outputs a degree of failure for each category for a plurality of components of the machine from the deep learning model unit 220.

Accordingly, the apparatus for diagnosing mechanical equipment failure using the auto encoder and the deep learning according to the present invention uses the auto-encoder 210 and the artificial neural network 200 in which the deep learning model is embedded to determine the type and cause of the failure of the components of the mechanical equipment. It has the advantage of being able to diagnose correctly.

3 is a schematic diagram showing an encoding process of the auto encoder 210 according to the present invention.

As shown in FIG. 3, the auto encoder 210 includes an input layer unit 211 in which the input data transferred from the input unit 100 is converted into an input layer, and the input layer is encoded to have a hidden feature map. The layer may include a hidden layer unit 212 compressed to a layer, and an output layer unit 213 to which the hidden layer is decoded and reconstructed as an output layer.

In more detail, the autoencoder 210 may represent a hidden layer having a feature map obtained by encoding the input data in a low dimension using the hidden layer, and more specifically, expressing the feature map using the output layer. Various expansions can be used to represent the decoded output layer.

That is, the auto encoder 210 compresses the input layer into a hidden layer, and the deep learning model unit 220 learns for efficient classification. At this time, the deep learning model unit 220 is the input layer and the output layer is the same, it can be said that the characteristics of the input data is compressed because it utilizes a low-dimensional node configured in the feature map of the hidden layer.

, 목표(target) 데이터를 벡터

라고 할 때

를 쌍으로 정할 수 있으며 입력레이어와 출력레이어의 쌍을 충분하게 확보해야 한다.Meanwhile, the deep learning model unit 220 needs a pair of an input layer and an output layer to learn the feature map, and the input layer is a vector.

, Vector the target data

When you say

Can be set as a pair, and sufficient pair of input layer and output layer should be secured.

Meanwhile, the auto encoder 210 performs the deep learning model unit 220 by calculating the mean of squared error so as to minimize a mean of squared error between the input layer and the output layer. Can be controlled. The root mean squared deviation is a commonly used measure of the difference between the estimated value or the value predicted by the model and the value observed in the real environment.

4 is a block diagram illustrating an example of an apparatus for diagnosing a failure of mechanical equipment using an auto encoder and deep learning according to the present invention.

As shown in FIG. 4, the input data may include vibration frequencies and vibration spectra generated in multiple parts of the mechanical equipment.

In this case, the apparatus for diagnosing a failure of the mechanical equipment using the auto encoder and the deep learning may further include a band pass filter 300 and an envelope 400.

The bandpass filter 300 gradually filters the vibration spectrum specific section of the input data until a defect signal according to the fault diagnosis information appears in the input data.

The envelope 400 models the vibration spectrum change of the input data transferred from the band pass filter 300.

For example, when the band pass filter 300 and the envelope 400 are applied to diagnose a failure of a bearing, the band pass filter 300 and the envelope 400 may be used in the vibration spectrum, which is input data obtained from the bearing. Gradually filter out unwanted portions of the vibration spectrum until a defect signal appears. This process is called enveloping.

FIG. 5 is a graph illustrating an example of enveloping an apparatus for diagnosing a failure of a mechanical device using an auto encoder and deep learning according to the present invention in three dimensions.

Referring to FIG. 5, a three-dimensional change in envelope over time may be defined. In the deep learning model unit, data for learning the feature map may be limited to amplitude values at equal intervals within a range of a frequency axis. have.

FIG. 6 is a schematic diagram illustrating a failure diagnosis process of an acceleration machine in a state in which a vibration frequency and a vibration spectrum are input to input data of a device failure diagnosis apparatus using an auto-encoder and deep learning according to the present invention.

As shown in FIG. 6, the artificial neural network 200 has a probability value calculated by the output unit 230 as a softmax regression function for calculating a degree of failure for each category of the failure diagnosis failure information (Failure Impact Tab of FIG. 6). ) May be output as an output value, and a calibration unit 240 may be set to correct an error between the output value and the target value by gradient descent by setting a target value of the output value.

On the other hand, examples of the effect of the failure is an inner ring defect, outer ring defect, unbalance, mechanical looseness, misalignment, etc. of the bearing, the feature learning and failure classification pre-learning of the feature map in the deep learning model unit 220 is finished After that, new input data can be automatically classified.

FIG. 7 is a schematic diagram illustrating a failure diagnosis process of an acceleration machine in a state in which data generated from a plurality of sensors is input to input data of a device failure diagnosis apparatus using an auto-encoder and deep learning.

As shown in FIG. 7, the input data is data generated from a plurality of sensors separately installed on a plurality of components of a mechanical device. The input data is data of an acceleration sensor, data of a noise sensor, data of a light sensor, and data of a gyro sensor. , At least one of data of a humidity sensor, data of a magnetic sensor, data of an input sensor, data of a pressure sensor, and data of a temperature sensor.

The present invention is not limited to the above-described embodiments, and the scope of application is not limited, and various modifications can be made without departing from the gist of the present invention as claimed in the claims.

100: input unit
200: artificial neural network
210: Auto Encoder
211 input layer
212: hidden layer
213: output layer
220: deep learning model
230: output unit
240: correction unit
300: band pass filter
400: Envelope

Claims (8)

An input unit 100 for acquiring and storing a plurality of input data including a vibration frequency and a vibration spectrum generated from a plurality of components of a mechanical device for rotating or linear motion;
A band pass filter 300 which gradually filters a vibration spectrum specific section of the input data until a defect signal according to the fault diagnosis information appears in the input data;
An envelope (400) for modeling a change in vibration spectrum of input data transferred from the band pass filter (300);
Auto encoder 210 for converting a plurality of input data transferred from the input unit 100 into a feature map, and feature map and fault diagnosis limited to predetermined amplitude values at equal intervals within a range of a predetermined frequency axis among the data. The deep learning model unit 220, which has been trained in advance with information, and the output unit 230 for inputting the feature map to the deep learning model unit 220 and outputting a degree of failure per category for a plurality of parts of the mechanical equipment. Artificial neural network (200) comprising;
The auto encoder 210 is
An input layer unit 211 in which the input data transferred from the input unit 100 is converted into an input layer, a hidden layer unit 212 which is encoded into a hidden layer having a feature map encoded therein, and the hidden layer An output layer unit 213 is decoded and reconstructed as an output layer,
The hidden layer unit 212 is made of a hidden layer of less than four layers,
An apparatus for diagnosing mechanical equipment failure using an auto-encoder and deep learning, characterized in that the calculation and control are performed to minimize a mean of squared error between the input layer and the output layer.
delete delete delete delete According to claim 1, wherein the artificial neural network 200
The output unit 230 outputs a probability value calculated by the softmax regression function as the failure confidence level as an output value,
And a calibration unit (240) for setting a target value of the output value and correcting an error between the output value and the target value by a gradient descent method.
The method of claim 1, wherein the input data is
Device failure diagnosis device using auto-encoder and deep learning, which are data generated from multiple sensors separately installed in multiple parts of mechanical equipment.
The method of claim 7, wherein the input data is
It includes at least one of acceleration sensor data, noise sensor data, light sensor data, gyro sensor data, humidity sensor data, magnetic sensor data, input sensor data, pressure sensor data and temperature sensor data. Mechanical equipment failure diagnosis device using auto encoder and deep learning.

Images