JP7349339B2 - Abnormality diagnosis device and abnormality diagnosis method for vibrating machines - Google Patents

Description

The present invention relates to an abnormality diagnosing device and an abnormality diagnosing method for diagnosing an abnormality based on vibrations and sounds detected from a vibrating machine.

In recent years, there has been active development of abnormality diagnosis technology for large rotating machines such as wind power generators, with the aim of reducing downtime due to failures of mechanical elements.

For example, as a technique for diagnosing abnormalities in rolling bearings, an abnormality diagnosing method as disclosed in Patent Document 1 has been proposed. That is, when a scratch occurs on a component that constitutes a bearing (for example, an outer ring, an inner ring, a rolling element, a cage), the scratch comes into contact with the component due to rotation, and vibrations and noise are generated. This is detected by an acceleration sensor or microphone, and the detected waveform is processed as a signal. Furthermore, the value obtained through the processing is compared with a reference value to determine the presence or absence of an abnormality.

As signal processing, a method has been proposed in which frequency analysis is performed after envelope processing of a detected waveform. The vibration frequency (characteristic frequency) excited by the scratch can be determined geometrically, and the value of the spectrum at that frequency is recorded as the characteristic amount. Since the characteristic frequency differs depending on the location (component) where the scratch occurs, in Patent Document 1, the characteristic amount at each frequency is recorded. As a result, when a flaw occurs in any component, the feature amount at that characteristic frequency increases, making it possible to specify the location where the flaw occurs.

Japanese Patent Application Publication No. 2001-21453

The abnormality diagnosis method described in Patent Document 1 requires bearing dimensions and rotational speed to calculate the characteristic frequency, so it is necessary to change the values for each device and operating condition. In addition, as a feature quantity, it may be necessary to make a multifaceted judgment by combining not only the spectral value at the characteristic frequency but also statistical quantities such as effective value, deviation value, skewness, kurtosis, and crest factor. Knowledge is required.

Therefore, in order to use the abnormality diagnosis method of Patent Document 1, a high degree of specialized knowledge is required, and it is difficult for those who do not have such specialized knowledge to use it.

Therefore, the purpose of the present invention is to automatically extract features using machine learning when performing signal processing based on vibration and sound detection waveforms obtained from vibration machines. To provide an abnormality diagnosis device for a vibrating machine that allows even a person without knowledge to easily diagnose an abnormality in the vibrating machine.

In order to solve the above problems, an abnormality diagnosis device of the present invention is an abnormality diagnosis device that diagnoses abnormality of a vibrating machine based on the output of a vibration detection sensor, and generates image data based on the output of the vibration detection sensor. an imaging unit that performs machine learning on an abnormality detection model of the vibration machine using the image data, and extracts the output of the vibration detection sensor from the imaged input image using the machine learned abnormality detection model. An output image is generated by dimensionally compressing the feature amount, calculating a latent variable, and restoring using the latent variable , and detecting an abnormality in the vibrating machine by comparing the input image and the output image. An abnormality detection section.

According to the abnormality diagnosis device of the present invention, when performing signal processing based on detected vibration and sound waveforms obtained from a vibrating machine, feature values can be automatically extracted by machine learning. Even a person without advanced knowledge can easily diagnose an abnormality in a vibrating machine.

Problems, configurations, and effects other than those described above will be made clear by the description of the following examples.

Schematic diagram of an abnormality diagnosis device according to Example 1 and a large rotating machine to be diagnosed Flowchart diagram representing machine learning processing of the abnormality diagnosis device of Example 1 Flowchart diagram showing diagnostic processing of the abnormality diagnostic device of Example 1 Flowchart diagram showing abnormality detection processing of the abnormality diagnosis device of Example 1 Conceptual diagram showing abnormality detection processing of the abnormality diagnosis device of Example 2

Embodiments of the invention will be described below with reference to the drawings.

An abnormality diagnosis device 1 according to a first embodiment of the present invention will be explained using FIGS. 1 to 4.

FIG. 1 is a schematic diagram of an abnormality diagnosis apparatus 1 according to the present embodiment and a large rotating machine 3 that is a subject of the diagnosis. In FIG. 1, a large rotating machine 3 is shown as an example of a diagnostic target, but a vibrating machine without a rotating mechanism may also be a vibrating machine that generates vibrations during operation.

The abnormality diagnosis device 1 is a device that diagnoses abnormalities in a large rotating machine 3 based on the output of the vibration detection sensor 2, and as shown in FIG. 13. Note that the abnormality diagnosis device 1 is specifically a computer equipped with hardware such as an arithmetic unit such as a CPU, a main storage device such as a semiconductor memory, and an auxiliary storage device. The functions of the signal processing unit 11, the imaging unit 12, and the abnormality detection unit 13 are realized by the arithmetic unit executing the program loaded from the auxiliary storage device to the main storage device. The following describes well-known techniques in the computer field, omitting them as appropriate.

The vibration detection sensor 2 is, for example, an acceleration sensor installed near a vibration source of the large rotating machine 3 to detect acceleration, or a microphone installed near a noise source of the large rotating machine 3 to detect sound. The vibration detection sensor 2 may be any other type of sensor as long as it can detect vibrations or sounds, but in the following, the vibration detection sensor 2 is an acceleration sensor, and the output of the vibration detection sensor 2 is an acceleration signal. It will be explained as follows.

The large rotating machine 3 is, for example, a large rotating machine such as a wind power generator, and includes a rotating shaft 31, a bearing 32 that supports the rotating shaft 31, a gear 33 that accelerates or decelerates the rotating shaft 31, and a housing that accommodates these. It has a casing 34 and the like. Note that, in FIG. 1, the configuration on the input side is designated by the symbol a, and the configuration on the output side is designated by the symbol b.

<Machine learning processing>
First, the machine learning process of the model of the large rotating machine 3 by the abnormality diagnosis device 1 will be explained using FIG.

The machine learning process executed here analyzes a large amount of normal data collected from the large rotating machine 3 that has been confirmed to be normal, and causes the abnormality detection unit 13 to machine learn the feature amount of the normal data. This is a pre-processing that is essential for generating an abnormality detection model for diagnosing the large rotating machine 3.

First, in step S1, the vibration detection sensor 2 outputs a normal acceleration signal detected during normal operation of the large rotating machine 3 to the abnormality diagnosis device 1.

In step S2, the signal processing unit 11 of the abnormality diagnosis device 1 converts the analog acceleration signal output by the vibration detection sensor 2 into a digital acceleration signal that can be processed by the abnormality diagnosis device 1. In addition, in FIG. 1, the signal processing unit 11 is arranged in the abnormality diagnosis device 1, but the signal processing unit 11 is arranged in the vibration detection sensor 2, and a digital acceleration signal is input to the abnormality diagnosis device 1. It may also be used as a configuration.

Next, in step S3, the imaging unit 12 of the abnormality diagnosis device 1 images the digital acceleration signal. Specifically, this imaging process is a process of creating image data consisting of a graph displaying acceleration in time series, as shown on the right side of step S3. Then, the processing from steps S1 to S3 is repeated until the number of normal image data required for machine learning is accumulated. Although FIG. 1 illustrates a configuration including the signal processing section 11, the signal processing section 11 may be omitted if the imaging section 12 can process analog signals as they are.

Then, in step S4, the abnormality detection unit 13 performs machine learning on the large number of normal image data created in step S3, extracts the feature amount of the normal data, and creates an abnormality detection model of the normal large rotating machine 3. generate. Thereby, the abnormality detection unit 13 can diagnose abnormalities in the large rotating machine 3 by a method described later. Note that the model generation process executed in this step is, for example, "unsupervised learning" by a neural network executed by an autoencoder or the like. Through such machine learning, it is possible to extract feature amounts from normal image data and automatically adjust the weights of the encoder 13a and decoder 13b that define the abnormality detection model.

<Diagnostic processing>
Next, the diagnosis processing of the large rotating machine 3 by the abnormality diagnosis device 1 will be explained using FIGS. 3 and 4. Note that here, it is assumed that there is an abnormality in the large rotating machine 3 to be diagnosed.

The processing in steps S1 to S3 in FIG. 3 is equivalent to the processing in steps S1 to S3 in FIG. 2, but in FIG. In the process, image data with abnormal parts is generated.

In step 5, the abnormality detection unit 13 executes abnormality detection processing for the large rotating machine 3 based on the abnormality detection model generated in step S4 of FIG. 2 and the image data generated in step S3. Details of the process in step S5 will be explained using FIG. 4.

First, in step S51, the abnormality detection section 13 obtains, as the input image X, image data with an abnormal portion generated by the imaging section 12.

Next, in step S52, the abnormality detection unit 13 compresses the input image X having an abnormal portion using the encoder 13a whose weights have been adjusted by machine learning. The compression process executed here is a process of extracting a feature amount from the input image X having an abnormal portion, and then performing dimension compression of the extracted feature amount. Then, the anomaly detection unit 13 calculates the latent variable A based on the dimensionally compressed feature amount.

In step S53, the abnormality detection unit 13 restores the latent variable A calculated in step S52 using the decoder 13b whose weight has been adjusted by machine learning, and generates an output image Y.

In step S54, the abnormality detection unit 13 compares the input image X and the output image Y, and calculates the restoration error |X−Y|. During machine learning in FIG. 2, the weights of the encoder 13a and decoder 13b are automatically adjusted so that the restoration error is small when normal data is input. Through the feature amount restoration process, abnormal portions are removed from the input image X, and output data Y that can be regarded as normal data is generated. Therefore, the restoration error |X-Y| calculated in this step corresponds to the extracted abnormal part of the input image It is possible to detect the presence or absence of

After completing step S5, in step S6, the abnormality detection unit 13 calculates the degree of abnormality, which is an index indicating the degree of abnormality in the large rotating machine 3. This degree of abnormality is, for example, the ratio of the abnormal portion (restoration error) detected in step S54 to the input image X. If the degree of abnormality is higher than a predetermined reference value, the abnormality detection unit 13 3 can be diagnosed as abnormal.

As explained above, according to the abnormality diagnosis device 1 of the present embodiment, even an operator without specialized knowledge can generate an abnormality detection model by machine learning, so that the abnormality detection model can be used. This makes it easy to diagnose abnormalities in large rotating machines.

The abnormality detection unit 13 according to the second embodiment of the present invention will be explained using FIG. 5. Note that redundant explanation of common points with Example 1 will be omitted.

In this embodiment, as the input image X to the abnormality detection unit 13, an image obtained by frequency analysis of the detection waveform of the vibration detection sensor 2, or an image obtained by frequency analysis of the detection waveform of the vibration detection sensor 2 after envelope processing is used. In this way, since the amplitude of the waveform at each frequency is directly displayed on the input image X and the output image Y, abnormalities can be detected more easily than in the configuration of the first embodiment. Note that the input image It is also possible to create a composite image by adding together a plurality of arbitrary images from . By using such a composite image, the storage capacity can be reduced compared to the case where a storage area is secured for each type of image.

After generating the output image Y, the abnormal frequency can be extracted by calculating the restoration error |X−Y| between the input image X and the output image Y. With this, it is possible not only to calculate the degree of abnormality using a method similar to that in the first embodiment, but also to identify the cause of the abnormality based on the abnormal frequency.

Note that the present invention is not limited to the embodiments described above, and includes various modifications. For example, the embodiments described above are described in detail to explain the present invention in an easy-to-understand manner, and the present invention is not necessarily limited to having all the configurations described. Furthermore, it is possible to replace a part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment. Furthermore, it is possible to add, delete, or replace some of the configurations of each embodiment with other configurations.

1 Abnormality diagnosis device 11 Signal processing unit 12 Imaging unit 13 Abnormality detection unit 13a Encoder 13b Decoder 2 Vibration detection sensor 3 Large rotating machine 31 Rotating shaft 32 Bearing 33 Gear 34 Casing

Claims (6)

An abnormality diagnosis device for diagnosing an abnormality in a vibrating machine based on the output of a vibration detection sensor,
an imaging unit that generates image data based on the output of the vibration detection sensor;
Using the image data, machine learning is performed on an abnormality detection model of the vibration machine, and using the machine learned abnormality detection model, the feature amount extracted from the input image obtained by converting the output of the vibration detection sensor into an image is dimensionally compressed. , an abnormality detection unit that calculates a latent variable, generates an output image by restoring using the latent variable , and detects an abnormality of the vibrating machine by comparing the input image and the output image;
An abnormality diagnosis device for a vibrating machine characterized by having the following. The abnormality diagnosis device according to claim 1,
The anomaly detection model is
weights of an encoder that performs dimension compression on the feature quantities extracted from the input image and calculates latent variables;
weights of a decoder that generates the output image by restoring the latent variables;
An abnormality diagnosis device characterized in that the abnormality diagnosis device is defined by automatically adjusting by machine learning. The abnormality diagnosis device according to claim 1,
The abnormality diagnosing device is characterized in that the abnormality detection unit diagnoses the vibrating machine as abnormal if a restoration error between the input image and the output image is higher than a predetermined reference value. The abnormality diagnosis device according to any one of claims 1 to 3,
The input image includes image data that displays the output of the vibration detection sensor in time series, image data that displays the envelope waveform of the output of the vibration detection sensor in time series, and image data that frequency-analyzes the output of the vibration detection sensor. , or image data obtained by frequency analysis of the envelope waveform of the output of the vibration detection sensor, or composite image data obtained by adding a plurality of image data. The abnormality diagnosis device for a vibrating machine according to claim 4,
When the input image is a frequency-analyzed image, the abnormality detection unit calculates an abnormal frequency from a portion remaining after calculating a restoration error from the input image and the output image, and calculates an abnormal frequency based on the abnormal frequency. , an abnormality diagnosis device for a vibrating machine characterized by identifying the cause of the abnormality. An abnormality diagnosis method for diagnosing an abnormality in a vibrating machine based on the output of a vibration detection sensor, the method comprising:
generating image data based on the output of the vibration detection sensor;
Machine learning an abnormality detection model of the vibrating machine using the image data,
Using a machine-learned anomaly detection model , dimensionally compress the features extracted from the input image that is the output of the vibration detection sensor, calculate latent variables, and restore using the latent variables to create an output image. generate ,
A method for diagnosing an abnormality in a vibrating machine, characterized in that an abnormality in the vibrating machine is detected by comparing the input image and the output image.

Images