KR20210006832A - Method and apparatus for machine fault diagnosis - Google Patents

Abstract

The present invention relates to a method and an apparatus for machine fault diagnosis, which are able to, in diagnosing the presence or absence of a fault in a machine, extract characteristics from measured signals of various types such as machine operation noise, vibration, and heat and define the types of faults, to secure learned data, to use a fault sorter learning based on the learned data, and to sort the normal status of the machine or a large number of abnormal faults. The present invention is able to diagnose at any time a fault of the relevant machine by using a fault sorter which re-learns unique characteristics of the signals extracted from the machine to be diagnosed by using the learned data which are renewed frequently or regularly. According to the present invention, the method and the apparatus for the machine fault diagnosis are able to regularly and rapidly monitor and diagnose the normal status or faulty status of the machine by using the fault sorter which is automatically renewed through multiple signal measurements and analyses, to allow the operation environment of the machine to respond to changing situations, and to reflect new fault-type information.

Description

Machine fault diagnosis method and apparatus {Method and apparatus for machine fault diagnosis}

The present invention relates to machine failure diagnosis, and more particularly, to a method and apparatus for classifying defects in a machine device by using machine learning in diagnosing a failure of a machine device.

In the event of a malfunction or defect in a machine operating in real life and industrial sites, serious human injury and enormous economic loss may occur. Accordingly, fault diagnosis for detecting abnormal fault conditions of the machine early by accurately diagnosing various faults occurring in the machine has become a very important issue.

In a process in which numerous parts constituting a machine are connected to each other and driven, for example, various noises, vibrations, and heat generation occur due to gaps, interference, or friction between parts. Therefore, sound, vibration, temperature, current, etc. generated during machine operation are known as comprehensive signals representing the overall condition of the machine. That is, when the machine is aged or put into a defective state due to an external impact or environmental influence, the signals are used to determine whether the machine is in an abnormal defect state because the machine deviates from the characteristic pattern indicated in the normal state.

In the conventional fault diagnosis method according to the regular inspection cycle, the operator or the machine manager must directly perform signal measurement and analysis at certain time intervals (e.g., every hour, daily, weekly, monthly, or yearly). Not only is the utilization rate temporarily lowered, but the cost required for management continues to increase. On the other hand, recently, in order to perform fault diagnosis on machines placed in remote and inaccessible areas at all times, through the introduction of a sensor network, the condition of the machine is quickly identified through automated signal measurement to quickly determine whether there is a defect. Several ways to do this are being sought.

In particular, in order to remove noise from the raw signal to be measured and to effectively calculate a feature (feature or signature) that is meaningful for machine fault diagnosis, Fourier transform or wavelet transform as well as statistical value analysis in the time and frequency dimensions Feature extraction methods such as (Wavelet transform) are applied. The extracted feature may be used to determine whether a failure has occurred by comparing it with a threshold value according to a related standard of a corresponding machine or a rule set by a domain expert. However, such a pre-rule-based feature comparison method has a problem in that various machines are driven under their own operating conditions, and the surrounding environment in which the machines are operated is insufficient to cope with the reality situation that gradually changes over time.

Recently, a fault classifier created based on learning models such as machine learning, especially deep learning, to generate feature vectors from acoustic signals that are continuously collected from multiple sensors and to classify fault types of machines effectively. ) Has been proposed for fault diagnosis. In order to build a learning model-based defect classifier, first of all, it is necessary to secure enough learning data suitable for the operating environment of the machine to be diagnosed. In addition, if the environmental requirements of the machine are changed, the collected data is collected in the changed operating environment. It is necessary to reflect the features extracted from recent signals back into the existing classifier. For example, the operating noise signal of a particular machine may have unique characteristics that are different from other machines of the same kind due to environmental requirements, so the classifier for fault diagnosis of a particular machine needs to be continuously updated through relearning. There is.

However, such conventional learning model-based failure diagnosis methods are at the level of repetitively using a classifier generated only with feature vectors of signals and fixed defect type information acquired in advance under limited machine operation conditions. On the other hand, in order to accurately diagnose faults, it is desirable to measure various and large amounts of signals from various sensors around the machine, but at the current technology level, the computational cost required to measure and analyze acoustic signals and vibration signals and It is worth considering that it takes a lot of time. Therefore, in the case of a wireless sensor node having a power supply limitation, it is set to perform signal measurement periodically rather than constantly so as to maintain available power by saving as much power as possible for signal measurement and transmission. Accordingly, there is a need for a method of performing accurate and constant diagnosis while using a small number of sensors.

The present invention proposes a method and apparatus for solving the above-described problems in which it is difficult to cope with changes in the environment around a machine and a new type of defect diagnosis is impossible.

In order to solve the above problems, the present invention secures learning data by performing feature extraction and definition of defect types from various types of measurement signals such as machine operation noise, vibration, and heat in diagnosing the presence or absence of a malfunction of a mechanical device. It provides a method and apparatus for classifying a normal state of a machine or a plurality of abnormal defects using a learned defect classifier.

The present invention will be described in detail. The method/device for diagnosing a machine failure according to the present invention is largely composed of a regular diagnosis and a regular (or occasional) diagnosis. The constant diagnosis includes the steps/means for generating a feature vector from an acoustic signal measured from a target machine; A step/means for diagnosing a failure of the machine with respect to the generated feature vector using a previously learned defect classifier; And storing a newly generated feature vector and a fault diagnosis result as learning data. The periodic diagnosis includes the steps/means of extracting features of the multiple signals measured regularly (frequently) from a target machine and defining a defect type at a corresponding time point; A step/means for updating (re-learning) the defect classifier using the defined defect type, the feature vector and the failure diagnosis result stored during the regular diagnosis process as additional learning data; It includes a step/means for frequently or regularly performing fault diagnosis of the machine using the fault classifier retrained in this way.

In this way, it is possible to constantly diagnose whether the machine has a failure using a defect classifier that is re-learned with learning data updated from time to time or regularly with various unique signal characteristics and defect type information extracted from the machine to be diagnosed. .

The configuration and operation of the present invention introduced above will become more apparent through specific embodiments described with reference to the drawings.

The present invention can be applied to constant failure diagnosis of machines used in various fields such as construction machines including bearings, gears, chains, machine tools, refrigeration and air conditioning machines, elevators, semiconductor equipment, and industrial robots. In other words, in order to constantly and quickly monitor and diagnose the normal or fault condition of the machine, a fault classifier that is automatically updated through multi-signal measurement and analysis is used to respond to changes in the operating environment of the machine and reflect new fault type information. To be able to do it.

In addition, it can be used for automatic detection of machine defects in the Internet of Things environment and regular monitoring service of production automation facilities. In addition, it is possible to obtain an additional effect of constantly diagnosing faults with only sound signals in diagnosing the driving status of various machines and detecting defects. have.

1 is an explanatory diagram summarizing the concept of a method for diagnosing a machine failure according to the present invention
2 is a schematic configuration diagram of a method for diagnosing a machine failure according to the present invention
3 is a detailed process flow diagram of a method for diagnosing a machine failure according to the present invention
4 is a diagram illustrating a process of extracting a frequency dimension feature from an acoustic signal
5 is an exemplary view of data stored in a learning data storage/management unit during the regular diagnosis process of FIG. 3
6 is another exemplary view of data stored in the learning data storage/management unit in the constant diagnosis process of FIG. 3
7 is an exemplary diagram of a signal measured through step 400 of the periodic diagnosis process of FIG. 3
8 is an exemplary diagram of a defect type defined through step 600 of the periodic diagnosis process of FIG. 3
FIG. 9 is an exemplary view of data including a defined defect type stored in a learning data storage/management unit in the constant diagnosis process of FIG. 3

Advantages and features of the present invention, and a method of achieving them will become apparent with reference to the embodiments described in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but will be implemented in various forms different from each other, only this embodiment makes the disclosure of the present invention complete, and those skilled in the art to which the present invention pertains. It is provided to fully inform the scope of the invention to the person, and the invention is defined by the description of the claims.

On the other hand, terms used in the present specification are for explaining examples and are not intended to limit the present invention. In this specification, the singular form also includes the plural form unless specifically stated in the phrase. As used in the specification, "comprise" or "comprising" means the presence of one or more other components, steps, operations and/or elements other than the mentioned elements, steps, operations and/or elements, or Does not rule out addition.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In adding reference numerals to elements of each drawing, even though they are indicated on different drawings, the same elements are assigned the same reference numerals as much as possible, and in describing the present invention, a detailed description of related known configurations or functions If the gist of the present invention may be obscured, a detailed description thereof will be omitted.

1 is a conceptual explanatory diagram of a method for diagnosing a machine failure according to the present invention.

A feature vector is generated from the acoustic signal measured from the target machine (12), and a malfunction of the machine is diagnosed with respect to the generated feature vector using a previously learned defect classifier (14). In this process, the newly created feature vector and the fault diagnosis result are stored (16). This diagnosis process will be referred to as'continuous diagnosis' 10 in the present invention.

On the other hand, the characteristics of the multiple signals measured regularly (frequently) from the machine are extracted and the type of defect at the time is defined (18). The defect classifier is updated (retrained) by using the defect type defined as described above, the feature vector stored during the regular diagnosis 10, and the failure diagnosis result as additional learning data (20). Using this re-learned fault classifier, machine fault diagnosis is performed frequently or regularly. This diagnosis process will be referred to as a'periodic diagnosis' (or occasional diagnosis) 22 in the present invention.

The regular diagnosis 10 presupposes that the diagnosis is performed at relatively short time intervals, that is, more frequently than the regular diagnosis 22. A function module related to the flow of the regular diagnosis 10 and a function module performing the flow of the periodic diagnosis 22 may be implemented as separate devices and interlocked. For example, a device at the sensor node or gateway level of the sensor network processes the continuous diagnosis (10) flow, and the function of extracting features from multiple signals in the periodic diagnosis (22) flow and defining the defect type is a PC server level device. It can be implemented to be processed in

2 schematically illustrates the configuration of a method for diagnosing a machine failure according to the present invention, based on the concept of the above-described failure diagnosis. First, it is assumed that the defect classifier for diagnosing a machine defect condition is learned in advance by learning data separately prepared in advance (S05).

S10 is a step of generating a feature vector by acquiring an acoustic signal by a sensor installed in a machine to be diagnosed, extracting a feature in a frequency dimension. S20 is a step of diagnosing a normal state or an abnormal defect state of the machine by using the defect classifier and the feature vector learned in advance using the learning data. S30 is a step of combining the machine state information according to the fault diagnosis with the feature vector and storing it in the learning data. Finally, S40 is a step of extracting features from the multi-sensor signals measured by the target machine, generating defect type information by combining them, and updating the learning data. At this time, the defect classifier is relearned from time to time (or periodically) using the updated learning data (S25).

3 is for explaining a processing flow embodied in order to implement the concept of machine failure diagnosis described above. In FIG. 3, the solid arrow indicates the flow of the regular diagnosis 10, and the dotted arrow indicates the flow of the regular diagnosis 22. In FIG.

First, the flow of performing the regular diagnosis 10 will be described.

100: Acquisition of sound signals

First, in order to measure the operating noise of the machine 50 to be diagnosed, an acoustic signal measured around the machine being driven is acquired. To do this, a microphone is installed around the machine to collect sound signals for a certain period of time. In this case, the sampling rate and collection time of the acoustic signal to be collected may be adjusted according to the technical standard and operation period of the machine to be diagnosed. That is, in the case of rotating machines including electric motors, the sampling rate must be set so that at least one data point for one rotation can be secured in consideration of the number of revolutions per minute (RPM). For example, in the case of an electric motor operating at 2000RPM, it takes 0.03 seconds per rotation, so if you measure at a 20kHz sampling rate, 0.03 (second) × 20,000 (times/second) per rotation = 600 data points in total. Can be collected.

200: sound signal feature extraction

Next, a feature of a frequency dimension is extracted from the measured acoustic signal. To this end, the acoustic signal is first converted into numerical values for each frequency component through Fourier transform or wavelet transform. Here, the Fourier transform converts a signal value sampled in a time domain into an amplitude and a phase for each frequency component in a frequency domain. Further, the wavelet transform converts a signal value in a time domain into approximation coefficients for a low frequency component in a frequency domain and detail coefficients for a high frequency component. Next, from the numerical values converted to the frequency domain, a pattern that is easy to compare the normal operation state and the defect state of the machine is extracted as a feature. More specifically, among frequency components, components that are unnecessary for classifying a machine state or other noise-related components may be filtered through a band-pass filter. In addition, through data discretization or quantization, the size of neighboring frequency components is divided into a certain section or several points, then grouped and merged into representative values (e.g., average, median, maximum, minimum, etc.). I can.

4 shows a result of extracting a frequency dimension feature from an acoustic signal collected in the time domain through Fourier transform. The result of Fourier transform, band filter application, and discretization/quantization of the sound signal of a normal machine (a), and the result of Fourier transform, band filter application, and discretization/quantization of the sound signal of a faulty machine (b) are compared.

300: feature vector generation

A feature vector is generated from the extracted features. The feature vectors refer to values obtained by extracting features related to a normal or defective state, and are values matched to the input form of the defect classifier by vectorizing representative numerical values for each frequency component. The feature vector 310 is stored and managed 700 as learning data on the one hand, and provided to the defect classification step 800 on the one hand.

700: storage and management of training data, 800: defect classification

The generated feature vector 310 is stored and managed to be used as learning data for learning machine failure diagnosis (700). The stored and managed learning data 710 is data used for learning of the defect classifier in the defect classification step 800, and in the defect classification step 800, the defect classifier is learned in advance through the learning data 710 and continues. Proceed with learning.

900: fault diagnosis

Machine failure diagnosis is performed by determining whether a machine is defective from the feature vector 310 generated in step 300 by using the defect classifier learned in advance with the learning data 710. Here, as a learning model used in the defect classifier, SVM (Support Vector Machine), Random Forest, or Deep Neural Networks may be used.

In addition, the feature vector 310 used for the fault diagnosis and the fault diagnosis result 910 are stored in the learning data (700). As an example, FIG. 5 shows a situation in which a feature vector 310 generated from an acoustic signal at 10-second intervals and a fault diagnosis result 910 therefor are stored in the learning data storage/management unit 700.

As described above, the flow from the acoustic signal acquisition 100 to the fault diagnosis 900 may be constantly performed (constant diagnosis). As an example, FIG. 6 shows a situation in which the above constant diagnosis 10 flow is continuously performed every 10 seconds (refer to the Timestamp column). (The feature vector and the failure diagnosis result are shown in 300 according to the second embodiment of the present invention). This is an exemplary diagram explaining that the feature vector generated in and the fault diagnosis result derived from 900 are stored in 700.)

Next, a flow of performing the periodic diagnosis 22 will be described with reference to FIG. 3 again.

400: multiple signal measurements

Various types of sensor signals are measured frequently or regularly around the machine 50 to be diagnosed (400). Unlike the flow of the regular diagnosis 10, the periodic diagnosis 22 may be performed when a diagnosis request by the administrator is requested, or at a regular time point of less frequency than the regular diagnosis flow. For example, when a fault diagnosis result in the constant diagnosis flow is classified as an abnormal state more than a certain number of times, it may be performed to reconfirm the fault state of the corresponding machine 50. That is, as shown in FIG. 6, when the abnormal state is diagnosed twice or more as a result of the fault diagnosis, the multi-signal measurement 400 may be performed.

Here, the signal measurement may measure a signal from one or more homogeneous or heterogeneous sensors. For example, the sound signal measured in step 400 may be collected from a separate microphone different from the microphone used for measuring the sound signal in step 100. As an example, FIG. 7 shows a situation in which one or more sensor signals are measured for an acoustic signal for operating noise, a temperature value for heat generation, or an acceleration signal for mechanical vibration.

500: Feature extraction for each signal

Features are extracted for each signal from the measured multiple signals. Specifically, after removing noise or errors in signal data for each individual signal, time-frequency dimensional transformation and statistical representative values are calculated to extract unique features for each individual signal. In this case, in order to extract a feature of a vibration or an acoustic signal, feature extraction in the same method or a different method as in the frequency dimension feature extraction performed in the acoustic signal feature extraction step 200 may be performed. For example, although Fourier transform-based feature extraction is performed in step 200, a wavelet transform-based feature extraction may be separately performed in step 500.

600: Defining the type of defect

Defects for each signal are determined by checking whether a normal pattern deviates or exceeds a threshold value according to the characteristics of each signal, and then the combined defect type is defined. For example, in FIG. 8 (a) is a case where a defect is determined in one acoustic signal ('sound 3') and two vibration signals ('vibration 2','vibration 3'), and the total defect type Is defined as'S1V2T0', '120', or'Abnormal'. Meanwhile, (b) of FIG. 8 is a case in which all signals are determined to be normal, and the combined defect type is defined as'S0V0T0', '000', or'Normal'.

The defect type defined as described above is sent to the learning data storage/management step 700 to update the learning data. Specifically, the defect type for the feature vector from the most recent failure diagnosis result to a specific number or specific time point is updated. For example, FIG. 9 shows that the defect type defined in step 600 is updated with respect to the learning data (refer to FIG. 6) within the last 1 minute among feature vectors for which continuous diagnosis was performed every 10 seconds. That is, if the defect type through multi-signal measurement at 18:01:10 is defined as'S1V2T0', the defect for the feature vector within the last minute (18:0:10-18:01:10) Update the type information.

Thereafter, in the defect classification step 800, the defect classifier is retrained using the feature vector in which the defect type is defined as the learning data 710. For example, the fault classifier gradually updates the learning model so that a feature vector that was previously diagnosed as normal can be diagnosed as abnormal, or conversely, a feature vector that has been diagnosed as abnormal can be diagnosed as normal. At this time, the time point at which the defect classifier performs re-learning is when a certain amount of the feature vector defined with the defect type is added, and when a new defect type different from the existing defect type of the learning data is added, every periodic time interval. , Or may be determined by the administrator's request.

Above, the configuration of the present invention has been described in detail through a preferred embodiment of the present invention, but those of ordinary skill in the art to which the present invention pertains, the present invention is disclosed in the present specification without changing the technical spirit or essential features. It will be appreciated that it may be implemented in a specific form different from that of. It should be understood that the embodiments described above are illustrative in all respects and not limiting. The scope of protection of the present invention is determined by the claims described later rather than the detailed description, and all changes or modifications derived from the scope of the claims and their equivalent concepts should be interpreted as being included in the technical scope of the present invention. .

Claims (1)

Generating a feature vector from the acoustic signal measured from the target machine;
Diagnosing a failure of the machine with respect to the generated feature vector using a previously learned defect classifier;
Storing newly generated feature vectors and fault diagnosis results as learning data;
Extracting features of the multi-signal measured from the target machine and defining a defect type at a corresponding time point;
Relearning the defect classifier by using the defined defect type, the feature vector stored during the regular diagnosis process, and the failure diagnosis result as additional learning data;
And performing a fault diagnosis of the machine using the retrained fault classifier.

Images