JP2021140403A - Processing method, diagnosis method of rotary machine, computer program, learning model production method, and diagnosis device - Google Patents

Abstract

To provide a diagnosis method of a rotary machine capable of highly precisely determining a sort of abnormality of the rotary machine, a computer program, a learning model production method, and a diagnosis device.SOLUTION: A diagnosis method of a rotary machine includes: acquiring time-sequential acceleration data time-sequentially recording an acceleration of a vibration measured by a vibration sensor attached to the rotary machine; producing plural sorts of feature data items from the time-sequential acceleration data; inputting the plural sorts of data items to a first learning model that outputs state data representing the state of the vibration when plural sorts of data items out of the time-sequential acceleration data and the plural sorts of feature data items are input; and acquiring the state data output by the first learning model.SELECTED DRAWING: Figure 1

Description

The present invention relates to a rotary machine diagnostic method, a computer program, a learning model generation method, and a diagnostic device for diagnosing the state of the rotary machine from the measured vibration of the rotary machine.

Currently, many rotating machines are used in various devices. The rotating machine is, for example, an electric motor, a turbine, an internal combustion engine, a blower, a pump, or the like. When an abnormality such as improper fixing, misalignment, or breakage occurs in the rotating machine, the vibration generated in the rotating machine changes as compared with the normal state. By measuring the vibration generated in the rotating machine and analyzing the vibration, it is diagnosed whether the rotating machine is normal or abnormal. For example, a person is diagnosed based on the measured vibration. In addition, a technique for performing diagnosis by a computer has also been developed. Patent Document 1 discloses a technique of acquiring a vibration waveform with a vibration sensor attached to a rotating machine, calculating a feature amount of the vibration waveform, and diagnosing a bearing of the rotating machine from the feature amount using a deep learning model. There is. Further, Patent Document 2 discloses a technique for diagnosing the root cause of an abnormality of a turbomachine by a fuzzy logic-based engine provided with a physical model.

Japanese Unexamined Patent Publication No. 2019-132809 Japanese Unexamined Patent Publication No. 2012-180831

Since a plurality of types of abnormalities can occur in a rotating machine, it is desirable to determine the type of abnormality that has occurred in the diagnosis of the rotating machine. At present, there is a limit to automatically determining the type of abnormality that has occurred in computer-based diagnosis. In the human diagnosis of the rotating machine, a comprehensive diagnosis is made using a plurality of types of information such as visual appearance information and vibration information by hearing and tentacles. Even in computer-based diagnosis, a technique capable of diagnosing a rotating machine with higher accuracy is desired by following a human-based diagnostic method.

The present invention has been made in view of such circumstances, and an object of the present invention is a diagnostic method for a rotating machine, a computer program, and a learning model generation capable of accurately determining the type of abnormality of the rotating machine. The purpose is to provide a method and a diagnostic device.

The method for diagnosing a rotating machine according to the present invention is a method for diagnosing a rotating machine by acquiring time-series acceleration data in which the acceleration of vibration measured by a vibration sensor attached to the rotating machine is recorded in time series. A plurality of types of feature data are generated from the time-series acceleration data, and when a plurality of types of data among the time-series acceleration data and the plurality of types of feature data are input, state data representing the vibration state is output. It is characterized in that the plurality of types of data are input to the first learning model to be used, and the state data output by the first learning model is acquired.

The method for diagnosing a rotating machine according to the present invention is a method for diagnosing a rotating machine by acquiring time-series acceleration data in which the acceleration of vibration measured by a vibration sensor attached to the rotating machine is recorded in time series. The first feature data in which the amount other than the acceleration representing the vibration is recorded in time series is acquired, the second feature data is generated from the time series acceleration data or the first feature data, and the time series acceleration data, the said. When a plurality of types of data among the first feature data and the second feature data are input, the plurality of types of data are input to the first learning model that outputs the state data representing the vibration state. It is characterized in that the state data output by the first learning model is acquired.

The computer program according to the present invention acquires time-series acceleration data in which the acceleration of vibration measured by a vibration sensor attached to a rotating machine is recorded in time-series, and obtains a plurality of types of feature data from the time-series acceleration data. The plurality of types of data are generated and output to a first learning model that outputs state data representing the state of vibration when a plurality of types of data among the time-series acceleration data and the plurality of types of feature data are input. Is input, and the computer is made to execute the process of acquiring the state data output by the first learning model.

In the computer program according to the present invention, the plurality of types of feature data include first feature data in which quantities other than acceleration representing the vibration are recorded in time series.

In the computer program according to the present invention, the first feature data is time-series velocity data in which the vibration velocity is recorded in time series, and time-series envelope data in which the envelope of the vibration acceleration is recorded in time series. , Or the process of generating the time-series velocity data by integrating the time-series acceleration data, which is the time-series displacement data in which the displacement of the vibration is recorded in time series, and the envelope is generated from the time-series acceleration data. By doing so, the computer is made to execute the process of generating the time-series envelope data or the process of generating the time-series displacement data by integrating the time-series acceleration data twice.

In the computer program according to the present invention, the value of the time-series acceleration data is included in the predetermined first range, the value of the time-series velocity data is included in the predetermined second range, or the above. When the value of the time-series displacement data is included in the predetermined third range, the rotating machine is characterized in that the computer further executes a process of determining that it is normal.

In the computer program according to the present invention, the plurality of types of feature data are the frequency spectrum obtained from the time-series acceleration data or the first feature data, or the amplitude obtained from the time-series acceleration data or the first feature data. It is characterized by including the second feature data which is a probability distribution.

In the computer program according to the present invention, the plurality of types of feature data include a frequency spectrum obtained from the time series acceleration data or a second feature data which is a probability distribution of amplitude obtained from the time series acceleration data. It is a feature.

In the computer program according to the present invention, the first learning model has a one-to-one correspondence with the plurality of types of data, and a plurality of convolutional neural networks that output a feature vector of the data when the corresponding data is input. It is characterized in that it includes a fully connected neural network that outputs the state data when a plurality of feature vectors output by the plurality of convolutional neural networks are input.

In the computer program according to the present invention, the first learning model includes a plurality of sets including a plurality of convolutional neural networks and one fully connected neural network, and a plurality of convolutional neural networks included in each set. The network has a one-to-one correspondence with a plurality of specific types of data among the plurality of types of data, and outputs a feature vector of the plurality of specific types of data when the corresponding specific types of data are input. The fully connected neural network included in each set is characterized in that the state data is output when a plurality of feature vectors output by a plurality of convolutional neural networks included in the same set are input.

The computer program according to the present invention acquires the attribute information of the rotating machine, and when the state data and the attribute information are input, the computer program outputs the factor information related to the factor of the vibration state to the second learning model. It is characterized in that a computer further executes a process of inputting data and the attribute information and acquiring the factor information output by the second learning model.

The computer program according to the present invention is characterized in that the computer further executes a process of outputting a diagnostic report of the rotating machine according to the content of the factor information.

The computer program according to the present invention outputs the content of the factor information, accepts the correction of the factor information, and causes the computer to further execute a process of re-learning the second learning model based on the received correction. It is characterized by.

The computer program according to the present invention acquires cases in which the degree of similarity with the diagnosis results for the rotating machine is within a predetermined range from a database that records a plurality of cases including past diagnosis results, and outputs the acquired cases. It is characterized in that the processing is further executed by a computer.

The learning model generation method according to the present invention includes time-series acceleration data in which the acceleration of vibration measured in the past by a vibration sensor attached to a rotating machine is recorded in time series, and a plurality of methods generated from the time-series acceleration data. A plurality of types of feature data are acquired, and for the time-series acceleration data and the plurality of types of feature data, state data representing the vibration state is acquired, and the time-series acceleration data and the plurality of types of feature data are acquired. When inputting time-series acceleration data in which the acceleration of an arbitrary vibration is recorded in a time-series manner and a plurality of types of feature data generated from the time-series acceleration data as training data, the arbitrary It is characterized in that a first learning model that outputs state data representing the state of vibration of is generated.

The learning model generation method according to the present invention obtains the time-series acceleration data by reading the time-series acceleration data from a database recording a plurality of cases including past diagnosis results without using a physical model. It is characterized in that the state data is acquired by acquiring and reading the state data from the database.

The learning model generation method according to the present invention includes state data representing the state of vibration measured in the past by a vibration sensor attached to the rotating machine, attribute information of the rotating machine, and factor information regarding factors of the state of vibration. Is acquired in a plurality of ways, and the state data representing the vibration state of an arbitrary rotating machine and the attribute information of the arbitrary rotating machine are input by using the plurality of ways of the state data, the attribute information and the factor information as training data. A second learning model that outputs factor information regarding the factors of the vibration state is generated.

The learning model generation method according to the present invention reads the state data, the attribute information, and the factor information from a database that records a plurality of cases including past diagnosis results, thereby reading the state data, the attribute information, and the above. It is characterized by acquiring factor information.

The diagnostic device according to the present invention is a diagnostic device for diagnosing a rotating machine, and includes an acceleration acquisition unit that acquires time-series acceleration data in which the acceleration of vibration measured by a vibration sensor attached to the rotating machine is recorded in time series. The state of vibration is represented when a generation unit that generates a plurality of types of feature data from the time-series acceleration data and a plurality of types of data among the time-series acceleration data and the plurality of types of feature data are input. It is characterized by including a first learning model that outputs state data and a state data acquisition unit that inputs the plurality of types of data and acquires the state data.

In one embodiment of the present invention, the rotating machine is diagnosed using not only the time-series acceleration data that records the acceleration of the vibration generated in the rotating machine but also the feature data generated from the time-series acceleration data. By using the learning model, state data representing the state of vibration can be easily obtained from a plurality of types of information regarding vibration. The state data clarifies what kind of state the vibration is in, such as a normal state or a state in which a specific frequency component is large. By using a plurality of types of information regarding vibration, it is possible to perform a diagnosis with higher accuracy than a conventional diagnosis.

In one embodiment of the present invention, not only the time-series acceleration data but also the first feature data in which quantities other than the acceleration representing vibration are recorded in time series, or the second feature data generated from the time-series acceleration data or the first feature data. The rotating machine is diagnosed using the feature data. By using the learning model, state data representing the state of vibration can be easily obtained from a plurality of types of information regarding vibration. The state data makes it clear what the vibration is like. By using a plurality of types of information regarding vibration, it is possible to perform a diagnosis with higher accuracy than a conventional diagnosis.

In one embodiment of the present invention, the feature data includes first feature data in which quantities other than acceleration representing vibration are recorded in time series. The rotating machine is diagnosed using not only the time-series acceleration data but also the first feature data. By using the learning model, state data representing the state of vibration can be easily obtained from a plurality of types of information regarding vibration. By using a plurality of types of information regarding vibration, it is possible to perform a diagnosis with higher accuracy than a conventional diagnosis.

In one embodiment of the present invention, the first feature data is time-series velocity data in which the vibration velocity is recorded in time series, time-series envelope data in which the envelope of vibration acceleration is recorded in time series, or vibration. It is time series displacement data which recorded the displacement of. Other first feature data can be easily generated from the time-series acceleration data obtained by the vibration sensor. By using the first feature data that represents the time change of the amount other than the acceleration of vibration in addition to the time series acceleration data, the rotating machine is diagnosed from a multimodal viewpoint using multiple types of information related to vibration. be able to.

In one embodiment of the present invention, the rotating machine is determined to be normal when the values of the time-series acceleration data, the time-series velocity data, or the time-series displacement data are included in the predetermined range. When the fluctuation of the acceleration, velocity or displacement of the vibration is small, it is possible to determine that the rotating machine is normal because the vibration having a large amplitude is not generated.

In one embodiment of the present invention, the feature data includes a second feature data. The second feature data is a frequency spectrum obtained from the time-series acceleration data or the first feature data, or a probability distribution of amplitude obtained from the time-series acceleration data or the first feature data. The frequency spectrum and the probability distribution of amplitude represent the characteristics of vibration acceleration, velocity, envelope or displacement. Further, by using the second feature data, it is possible to diagnose the rotating machine from a multimodal viewpoint using a plurality of types of information regarding vibration.

In one embodiment of the present invention, the first learning model outputs a plurality of convolutional neural networks that output feature vectors when corresponding data are input, and state data when a plurality of feature vectors are input. Includes a fully connected neural network. The plurality of convolutional neural networks obtains feature vectors that represent the features of the data according to the local correlation contained in the input data. Fully coupled neural networks provide state data representing a particular type of state with respect to vibration, depending on the characteristics of the plurality of types of data used.

In one embodiment of the invention, the first learning model comprises a plurality of sets of neural networks. The set includes a plurality of convolutional neural networks that output feature vectors when the corresponding data is input, and a fully connected neural network that outputs state data when a plurality of feature vectors are input. State data representing various quantities of state with respect to vibration can be obtained, depending on the characteristics of the plurality of types of data used.

In one embodiment of the present invention, the attribute information and the state data of the rotating machine are input to the second learning model, and the second learning model outputs the factor information regarding the factors of the vibration state. By using the second learning model, factor information can be easily obtained from the state data and the attribute information. From the factor information, it becomes clear that the state of vibration is caused by the state of the rotating machine.

In one embodiment of the present invention, a diagnostic report of the rotating machine according to the content of the factor information is output. The user can confirm the diagnosis result of the rotating machine from the diagnosis report.

In one embodiment of the present invention, the factor information is modified and the second learning model is relearned based on the modification of the factor information. The accuracy of diagnosis is improved by re-learning the second learning model.

In one embodiment of the present invention, a case similar to the obtained diagnostic result is output from a database recording a plurality of past cases. The user can confirm the content of the case similar to the diagnosis result, and can guess under what circumstances the acquired vibration is likely to occur.

In one embodiment of the present invention, learning is performed using the time-series acceleration data and feature data relating to the vibration measured in the past and the state data representing the state of the vibration measured in the past as training data, and the first method is performed. A training model is generated. By using such a first learning model, state data representing the state of vibration generated in the rotating machine can be obtained from the time series acceleration data and feature data.

In one embodiment of the present invention, learning is performed using state data representing the vibration state measured in the past, attribute information of the rotating machine whose vibration has been measured, and factor information related to the factors of the vibration state as training data. Then, the second learning model is generated. By using such a second learning model, factor information can be obtained from the state data and the attribute information, and the factor of the state of the vibration generated in the rotating machine can be clarified from the factor information.

In one embodiment of the present invention, a learning model is trained using data read from a database recording a plurality of past cases as training data. Without using a physical model, it is possible to generate a learning model that enables appropriate diagnosis by using the diagnosis results of a large number of rotating machines accumulated in the past by human diagnosis.

The present invention has excellent effects, such as being able to diagnose the rotating machine more accurately by performing the diagnosis of the rotating machine from a multimodal viewpoint using a plurality of types of information regarding vibration.

It is a schematic diagram which shows the structure of the diagnostic system which concerns on Embodiment 1 for diagnosing a rotating machine. It is a block diagram which shows the example of the functional structure inside the diagnostic apparatus. It is a block diagram which shows the example of the functional structure inside the learning device. It is a flowchart which shows the procedure of the process of the diagnosis of a rotating machine performed by a diagnostic apparatus. It is a conceptual diagram which shows the example of the process which generates the 1st feature data. It is a conceptual diagram which shows the example of the process which generates the 2nd feature data. It is a conceptual diagram which shows the outline of the structure and function of a vibration determination model and a diagnostic model. It is a conceptual diagram which shows an example of the structure and function of the set of NN included in the vibration determination model. It is a conceptual diagram which shows other example of the structure and function of the set of NN included in the vibration determination model. It is a conceptual diagram which shows the content example of the attribute information. The attribute information includes information indicating the structure of the rotating machine. It is a conceptual diagram which shows an example of the content of the factor information. It is a schematic diagram which shows the example of the image which the display part displays. It is a conceptual diagram which shows the content example of the case database. It is a flowchart which shows the procedure of the process which the learning apparatus generates a set of NN which consists of a plurality of CNNs and one FCNN included in a vibration determination model. It is a flowchart which shows the procedure of the process which a learning apparatus generates a diagnostic model. It is a flowchart which shows the procedure of the process of relearning performed by a learning apparatus. It is a schematic diagram which shows the structure of the diagnostic system which concerns on Embodiment 2. It is a block diagram which shows the functional configuration example inside the terminal apparatus.

Hereinafter, the present invention will be specifically described with reference to the drawings showing the embodiments thereof.
<Embodiment 1>
FIG. 1 is a schematic diagram showing a configuration of a diagnostic system according to the first embodiment for diagnosing a rotating machine. The rotating machine 2 is a machine having a rotating shaft and partially rotating. For example, the rotary machine 2 is an electric motor, a turbine, an internal combustion engine, a blower, a pump, or the like. The rotary machine 2 to be diagnosed is mainly a small and medium-sized rotary machine having a rated output of 75 kW or less shown in "Annex B of Japanese Industrial Standard JIS B0906 Mechanical Vibration". A vibration sensor 31 is attached to the rotating machine 2. The vibration sensor 31 measures the acceleration of vibration generated in the rotating machine 2 and outputs a signal indicating the value of the measured acceleration. For example, the vibration sensor 31 is attached to the bearing of the rotating machine 2. A plurality of vibration sensors 31 may be attached to the rotating machine 2. For example, a vibration sensor 31 that measures vibration in the horizontal direction and a vibration sensor 31 that measures vibration in the vertical direction are attached to one end of the rotating shaft of the rotating machine 2, and the other end of the rotating shaft is also attached to the horizontal and horizontal directions. A vibration sensor 31 for measuring vertical vibration is attached, and a total of four vibration sensors 31 are attached.

The vibration sensor 31 is connected to the A / D (analog-to-digital) converter 32 by wire or wirelessly. The A / D converter 32 receives the signal output by the vibration sensor 31, converts the received signal into a digital signal, and outputs the converted signal. That is, the A / D converter 32 outputs a signal that digitally indicates the value of the acceleration of vibration measured by the vibration sensor 31. When a plurality of vibration sensors 31 are attached to the rotating machine 2, the plurality of vibration sensors 31 are connected to the A / D converter 32. The A / D converter 32 receives a plurality of signals output by the plurality of vibration sensors 31, performs A / D conversion, and outputs a plurality of signals. The A / D converter 32 is connected to a diagnostic device 1 that diagnoses the rotary machine 2. The diagnostic device 1 receives the signal output by the A / D converter 32. That is, the diagnostic device 1 receives a signal indicating the value of the acceleration of vibration measured by the vibration sensor 31. The vibration sensor 31 and the A / D converter 32 may be integrated.

The diagnostic device 1 is connected to a communication network N such as a LAN (local area network) or the Internet. An output device 33 such as a printer is connected to the communication network N. Further, a learning device 4 for learning a learning model necessary for diagnosing the rotating machine is connected to the communication network N.

FIG. 2 is a block diagram showing an example of the internal functional configuration of the diagnostic device 1. The diagnostic device 1 is configured by using a computer such as a personal computer or a tablet computer. The diagnostic device 1 includes a calculation unit 11, a memory 12 for storing temporary data generated in association with the calculation, and a storage unit 13. The arithmetic unit 11 is configured by using, for example, a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), or a multi-core CPU. Further, the arithmetic unit 11 may be configured by using a quantum computer. The memory 12 is, for example, a RAM (Random Access Memory). The storage unit 13 is non-volatile, for example, a hard disk.

Further, the diagnostic device 1 includes a drive unit 14 that reads information from a recording medium 10 such as an optical disk, an operation unit 15 that receives an operation from a user, a display unit 16 that displays an image, an interface unit 17, and a communication unit. It has 18 and. The operation unit 15 receives information such as text by receiving an operation from the user. The operation unit 15 is, for example, a keyboard or a pointing device. The display unit 16 is, for example, a liquid crystal display or an EL display (Electroluminescent Display). The interface unit 17 inputs and outputs signals. An A / D converter 32 is connected to the interface unit 17. The communication unit 18 is connected to the communication network N and communicates with the outside of the diagnostic device 1. The communication unit 18 transmits / receives data to / from the output device 33 and the learning device 4 through the communication network N.

The calculation unit 11 causes the drive unit 14 to read the computer program 131 recorded on the recording medium 10, and stores the read computer program 131 in the storage unit 13. The calculation unit 11 executes the processing required for the diagnostic apparatus 1 according to the computer program 131. The computer program 131 may be downloaded from the outside of the diagnostic apparatus 1. In this case, the diagnostic device 1 does not have to include the drive unit 14.

The diagnostic device 1 receives the signal output by the A / D converter 32 at the interface unit 17. That is, the diagnostic device 1 receives a signal indicating the value of the acceleration of vibration measured by the vibration sensor 31. When diagnosing the rotating machine 2, the vibration sensor 31 repeatedly measures the acceleration of vibration, and the diagnostic device 1 intermittently receives a signal indicating the value of the acceleration of vibration. The value of the vibration acceleration may be an absolute value or a relative value from a predetermined reference. The storage unit 13 stores the time-series acceleration data 134 in which the acceleration of vibration is recorded in time series. The time-series acceleration data 134 is data in which the value of the vibration acceleration measured by the vibration sensor 31 is recorded in time series. Each time a signal from the A / D converter 32 is received, the calculation unit 11 performs a process of recording the values indicated by the signals in the time-series acceleration data 134 in the order of reception. When a plurality of vibration sensors 31 are attached to the rotating machine 2, the calculation unit 11 stores a plurality of time-series acceleration data 134 corresponding to the plurality of vibration sensors 31 in the storage unit 13. The vibration sensor 31 continues to measure vibration for a long time of at least 10 seconds to 150 seconds. The time-series acceleration data 134 records the time change of the acceleration of the vibration over a long period of at least 10 seconds to 150 seconds. Irregular vibration can be captured by obtaining data on long-term vibration of 10 seconds or more. By obtaining data on long-term vibrations up to 150 seconds, most long-period vibrations can be captured.

The diagnostic device 1 includes a vibration determination model 132, which is a learning model used to determine the vibration state from the time-series acceleration data 134. The vibration determination model 132 corresponds to the first learning model. As will be described later, the vibration determination model 132 is learned to output state data representing the vibration state when data indicating vibration such as time-series acceleration data 134 is input. Further, the diagnostic device 1 includes a diagnostic model 133, which is a learning model used for diagnosing the rotary machine 2 from the state data. The diagnostic model 133 corresponds to the second learning model. As will be described later, the diagnostic model 133 is learned to output factor information regarding the factors of the vibration state when state data is input.

The vibration determination model 132 and the diagnosis model 133 are realized by the arithmetic unit 11 executing information processing according to the computer program 131. The storage unit 13 stores data necessary for realizing the vibration determination model 132 and the diagnostic model 133. Further, the vibration determination model 132 or the diagnostic model 133 may be configured by using hardware. For example, the vibration determination model 132 or the diagnostic model 133 may be configured to include a processor and a memory for storing necessary programs and data. Further, the vibration determination model 132 or the diagnostic model 133 may be realized by using a quantum computer.

FIG. 3 is a block diagram showing an example of the internal functional configuration of the learning device 4. The learning device 4 is configured by using a computer such as a server device. The learning device 4 includes a calculation unit 41, a memory 42, a storage unit 43, a drive unit 44, and a communication unit 45. The calculation unit 41 is configured by using, for example, a CPU, GPU, or multi-core CPU. Further, the arithmetic unit 41 may be configured by using a quantum computer. The memory 42 is, for example, a RAM. The storage unit 43 is non-volatile, for example, a hard disk. The communication unit 45 is connected to the communication network N and communicates with the outside of the learning device 4. The communication unit 45 transmits / receives data to / from the diagnostic device 1 through the communication network N.

The calculation unit 41 causes the drive unit 44 to read the computer program 431 recorded on the recording medium 40, and stores the read computer program 431 in the storage unit 43. The calculation unit 41 executes the processing required for the learning device 4 according to the computer program 431. The computer program 431 may be downloaded from the outside of the learning device 4. In this case, the learning device 4 does not have to include the drive unit 44.

The storage unit 43 stores a case database 432 that records a plurality of cases in which the rotating machine is actually diagnosed. The case database 432 records various data acquired for a large number of rotating machines in the past in association with the results of diagnosis performed for each rotating machine. For example, the contents of the case database 432 are created based on a diagnostic chart that reports the results of diagnosis and countermeasures, which were created for a large number of rotating machines by human diagnosis in the past. The learning device 4 uses the data recorded in the case database 432 as training data to perform a process of learning a vibration determination model and a diagnostic model, which are learning models. The learning process will be described later.

The diagnostic device 1 executes the diagnostic method of the rotating machine 2. FIG. 4 is a flowchart showing a procedure for diagnosing the rotating machine performed by the diagnostic apparatus 1. Hereinafter, the step is abbreviated as S. The calculation unit 11 executes the following processing according to the computer program 131. The calculation unit 11 reads the time-series acceleration data 134 from the storage unit 13 (S101). For example, the calculation unit 11 reads the latest predetermined amount of data from the time-series acceleration data 134 stored in the storage unit 13. The data to be read includes data recording the time change of the acceleration of vibration over a time of at least 10 seconds to 150 seconds. The process of S101 corresponds to the acceleration acquisition unit. Next, the calculation unit 11 generates first feature data in which quantities other than acceleration representing vibration are recorded in time series (S102).

In S102, the calculation unit 11 generates a plurality of types of first feature data from the time series acceleration data 134. The first feature data of multiple types are time-series velocity data in which the vibration velocity is recorded in time series, time-series envelope data in which the envelope of vibration acceleration is recorded in time series, and time-series of vibration displacement. It is the time series displacement data recorded in a specific manner. The calculation unit 11 generates time-series velocity data by integrating the time-series acceleration data 134. The calculation unit 11 generates time-series envelope data by generating an envelope from the time-series acceleration data 134. For example, the calculation unit 11 passes the acceleration recorded in the time-series acceleration data 134 through a high-pass filter or a high-frequency bandpass filter defined by "Japanese Industrial Standard JIS B0906 mechanical vibration", and rectifies and smoothes the subsequent waveform. By doing or Hilbert transforming, an envelope is generated. The calculation unit 11 generates time-series displacement data by integrating the time-series acceleration data 134 twice. In this way, the diagnostic device 1 acquires a plurality of types of first feature data. The calculation unit 11 stores the first feature data in the memory 12 or the storage unit 13.

FIG. 5 is a conceptual diagram showing an example of processing for generating the first feature data. As shown in FIG. 5, time-series velocity data, time-series envelope data, and time-series displacement data are generated from the time-series acceleration data 134. The figure shows the time-series acceleration data 134, the time-series velocity data, the time-series envelope data, and the waveforms representing the time-varying changes in the acceleration, velocity, envelope, and displacement of the vibration represented by the time-series displacement data. The horizontal axis of the waveform shows time, and the vertical axis shows the values of vibration acceleration, velocity, acceleration on the envelope, and displacement at each time point. The time-series acceleration data 134, the time-series velocity data, the time-series envelope data, and the time-series displacement data have different waveforms from each other and show different characteristics from each other.

The calculation unit 11 is not limited to the one that generates all kinds of first feature data in S102. The calculation unit 11 may generate only the first feature data required for the subsequent processing. For example, the calculation unit 11 may generate only the time series speed data in S102.

Next, the calculation unit 11 includes the acceleration value included in the time-series acceleration data 134 within the predetermined first range, and the velocity value included in the time-series velocity data is included in the predetermined second range. It is determined whether or not it is contained in (S103). For example, the value of acceleration is represented by a positive or negative value, and when the absolute value of acceleration is equal to or less than a predetermined first threshold value, the calculation unit 11 determines that the value of acceleration is included in the first range. judge. For example, the velocity value is represented by a positive or negative value, and when the absolute value of the velocity is equal to or less than a predetermined second threshold value, the calculation unit 11 determines that the velocity value is included in the second range. judge. The first threshold value and the second threshold value are stored in the storage unit 13 in advance or are included in the computer program 131.

The acceleration value included in the time-series acceleration data 134 is included in the predetermined first range, the velocity value included in the time-series velocity data is included in the predetermined second range, and the time. When the displacement value included in the series displacement data is included in the predetermined third range (S103: YES), the calculation unit 11 determines that the rotating machine 2 is normal (S104). When an abnormality occurs in the rotating machine 2, vibration with a large amplitude often occurs. When the acceleration, speed and displacement of the vibration are small, it is possible to determine that the rotating machine 2 is normal because the vibration having a large amplitude is not generated. In S103, the calculation unit 11 may perform a process of determining that the rotating machine 2 is normal when the acceleration is included in the first range, and the speed is included in the second range. If so, the rotating machine 2 may perform a process of determining that it is normal, and if the displacement is included in the third range, the rotating machine 2 may perform a process of determining that it is normal. In S103, the calculation unit 11 may perform a process of determining that the rotating machine 2 is normal when the acceleration and the speed are included in the predetermined range, and the acceleration and the displacement are included in the predetermined range. If so, the rotating machine 2 may perform a process of determining that it is normal, and if the speed and displacement are included in a predetermined range, the rotating machine 2 may perform a process of determining that it is normal. ..

Some values of acceleration included in the time-series acceleration data 134 are not included in the first range, some values of velocity included in the time-series velocity data are not included in the second range, or time. When a part of the displacement values included in the series displacement data are not included in the third range (S103: NO), the calculation unit 11 generates the second feature data (S105). In S105, the calculation unit 11 generates a plurality of types of second feature data representing the features related to the frequency distribution and the probability distribution of the amplitude of the time series acceleration data 134 and each first feature data. Specifically, the calculation unit 11 generates a frequency spectrum and an amplitude probability distribution from the time-series acceleration data 134, the time-series velocity data, the time-series envelope data, and the time-series displacement data as the second feature data. ..

In S105, the calculation unit 11 performs a fast Fourier transform of the time-series acceleration data 134, the time-series velocity data, the time-series envelope data, and the time-series displacement data, thereby performing the acceleration frequency spectrum, the velocity frequency spectrum, and the envelope. Generates the frequency spectrum of and the frequency spectrum of the displacement. Further, the calculation unit 11 calculates the frequency of the amplitude included in the time-series acceleration data 134, the time-series velocity data, the time-series envelope data, and the time-series displacement data, and calculates the probability density from the frequency to obtain the acceleration amplitude. Probability distribution of, velocity amplitude probability distribution, envelope amplitude probability distribution, and displacement amplitude probability distribution are generated. The probability distribution of amplitude is a probability distribution showing irregular vibration shown in "Japanese Industrial Standard JIS B0153 Mechanical Vibration / Impact Terminology". In this way, the diagnostic apparatus 1 acquires a plurality of types of second feature data. The calculation unit 11 stores each of the second feature data in the memory 12 or the storage unit 13.

FIG. 6 is a conceptual diagram showing an example of processing for generating the second feature data. As shown in FIG. 6, the probability distribution of the frequency spectrum and the amplitude is calculated from each of the time-series acceleration data 134, the time-series velocity data, the time-series envelope data, and the time-series displacement data. The figure shows a graph showing the frequency spectrum of acceleration and the probability distribution of acceleration amplitude. The horizontal axis of the frequency spectrum of acceleration indicates frequency, and the vertical axis indicates amplitude. The horizontal axis of the probability distribution of the acceleration amplitude indicates the amplitude or normalized amplitude included in the time-series acceleration data 134, and the vertical axis indicates the probability density. The frequency spectrum and amplitude probability distributions represent the characteristics of time-series acceleration data 134, time-series velocity data, time-series envelope data, and time-series displacement data.

The calculation unit 11 is not limited to the one that generates the second feature data of all kinds in S105. The calculation unit 11 may generate only the second feature data required for the subsequent processing. For example, the calculation unit 11 does not generate a frequency spectrum and an amplitude probability distribution for the first feature data that is not generated in S102 among the time-series velocity data, the time-series envelope data, and the time-series displacement data. May be good. The first feature data and the second feature data are included in a plurality of types of feature data representing vibration features. The processing of S102 and S105 corresponds to the generation unit.

Next, the calculation unit 11 uses the vibration determination model 132 to generate state data representing the vibration state (S106). The vibration determination model 132 includes a plurality of sets of neural networks in which a plurality of convolutional neural networks and one fully connected neural network are set as one set. Hereinafter, the neural network is referred to as NN, the convolutional neural network is referred to as CNN, and the fully connected neural network is referred to as FCNN (Fully connected neural network).

FIG. 7 is a conceptual diagram showing an outline of the configurations and functions of the vibration determination model 132 and the diagnostic model 133. The vibration determination model 132 includes a plurality of sets 51, 52, ... A set of a plurality of CNNs and one FCNN. The contents of each set are different from each other. The number of CNNs contained in each set may be the same or different from each other. The plurality of CNNs included in each set are time-series acceleration data 134, time-series velocity data, time-series envelope data, time-series displacement data, second feature data of time-series acceleration data 134, and second of time-series velocity data. There is a one-to-one correspondence with a plurality of types of data among the feature data, the second feature data of the time-series envelope data, and the second feature data of the time-series displacement data. It is desirable that the plurality of CNNs and FCNNs included in each set are NNs having a plurality of intermediate layers.

FIG. 8 is a conceptual diagram showing an example of the configuration and function of the NN set included in the vibration determination model 132. The NN set 51 includes CNN511a, 511b and 511c and FCNN512. CNN511a corresponds to the time series acceleration data 134, CNN511b corresponds to the frequency spectrum of acceleration, and CNN511c corresponds to the probability distribution of acceleration amplitude. Corresponding data is input to each CNN. In the example shown in FIG. 8, the time-series acceleration data 134 is input to CNN511a, the frequency spectrum of acceleration is input to CNN511b, and the probability distribution of acceleration amplitude is input to CNN511c. When the data is input to CNN511a, 511b and 511c, data processing such as normalization may be performed. CNN511a, 511b and 511c each output a feature vector representing the feature of the input data.

For FCNN512, a plurality of feature vectors output by CNN511a, 511b and 511c are input. When the data is input to FCNN512, data processing such as normalization may be performed. FCNN512 outputs state data 513 showing the state of vibration. In the example shown in FIG. 8, FCNN512 outputs state data 513 indicating the state of vibration acceleration. The state data 513 includes the probability or certainty that the acceleration of vibration contains each specific component such as an aperiodic component, a periodic specific component A1, and a periodic specific component A2. It has been. Alternatively, the state data 513 includes the intensities of each particular component included in the acceleration of vibration. In CNN511a, 511b and 511c, and FCNN512, FCNN512 is the state of acceleration of vibration when time-series acceleration data 134 is input to CNN511a, the frequency spectrum of acceleration is input to CNN511b, and the probability distribution of acceleration amplitude is input to CNN511c. It has been learned in advance so as to output the state data 513 representing.

By using CNN, a feature vector representing the characteristics of the data according to the local correlation contained in the input data can be obtained. For example, short-period changes in the values included in the time-series acceleration data 134 are reflected in the feature vector. The FCNN to which a plurality of feature vectors are input provides state data representing a particular type of quantity of state with respect to vibration, depending on the features of the plurality of types of data used.

FIG. 9 is a conceptual diagram showing another example of the configuration and function of the NN set included in the vibration determination model 132. The NN set 52 includes CNN 521a, 521b and 521c and FCNN522. CNN521a corresponds to time series velocity data, CNN521b corresponds to the frequency spectrum of velocity, and CNN521c corresponds to the probability distribution of velocity amplitude. Each of the CNNs 521a, 521b and 521c is input with corresponding data and outputs a feature vector representing the characteristics of the input data.

For FCNN522, a plurality of feature vectors output by CNN521a, 521b and 521c are input. FCNN522 outputs state data 523 representing the state of vibration velocity. The state data 523 includes the probability or certainty that the vibration velocity includes each specific component such as a sine waveform component, a component similar to the sine waveform, a specific component B1, and a specific component B2. It has been. Alternatively, the state data 523 includes the intensities of each particular component contained in the velocity of vibration. For CNN521a, 521b and 521c and FCNN522, when time-series velocity data is input to CNN521a, velocity frequency spectrum is input to CNN521b, and velocity amplitude probability distribution is input to CNN521c, FCNN522 changes the state of vibration velocity. It has been learned in advance so as to output the represented state data 523.

Similar to the examples shown in FIGS. 8 and 9, the plurality of CNNs included in each of the NN sets included in the vibration determination model 132 have a one-to-one correspondence with the specific plurality of types of data. The plurality of CNNs and FCNNs included in each set are learned in advance so that the FCNN outputs a specific type of state data when the corresponding plurality of types of data are input to the plurality of CNNs. The plurality of state data output by the FCNN included in the plurality of sets represents different kinds of states related to vibration. For example, in addition to the state data 513 representing the state of vibration acceleration or the state data 523 representing the state of vibration velocity, state data representing the state of the velocity waveform or the state data representing the state of the envelope of acceleration is output. Will be done. It is desirable that the combination of a plurality of types of data corresponding to a plurality of CNNs included in each set is different for each set. The FCNN included in each set outputs the state data, so that the vibration determination model 132 outputs a plurality of state data.

In S106, the calculation unit 11 inputs the corresponding data to the plurality of CNNs included in each of the NN sets included in the vibration determination model 132, and acquires the plurality of state data output by the vibration determination model 132. In this way, in S106, the calculation unit 11 generates state data. The vibration determination model 132 may be a learning model having a configuration other than the configuration shown in FIG. 7 as long as it is trained to input a specific plurality of types of data and output state data. The process of S106 corresponds to the state data acquisition unit.

Next, the calculation unit 11 acquires the attribute information of the rotating machine 2 (S107). FIG. 10 is a conceptual diagram showing an example of the content of attribute information. The attribute information is information about the rotating machine 2, and includes information determined by the specifications of the rotating machine 2 and information determined by the environment in which the rotating machine 2 is used. The information determined by the specifications includes the type of the rotating machine 2, the support structure, the drive system, the mechanical rating, and the like. Types of rotary machine 2 include electric motors, turbines, internal combustion engines, blowers, pumps, and the like. The support structure of the rotating machine 2 includes cantilever, cantilever, and the like. The drive system includes a motor direct connection, a belt drive, an inverter drive, and the like. The mechanical rating includes the capacity of the rotating machine 2, the number of rotations, the type of bearing, the nominal number of the bearing, and the like. The information determined by the environment in which the rotary machine 2 is used includes information on the gantry or foundation on which the rotary machine 2 is installed. Information on the gantry / foundation includes, for example, the type of material such as concrete or H-steel, the rigidity and flexibility, the presence or absence of cracks, and the presence or absence of deformation.

In the example shown in FIG. 10, information indicating a model (type of rotating machine 2) is included, and information specifying whether the model is an electric motor, a pump, or the like is included in the attribute information. For example, the attribute information includes information in which the corresponding model is represented by 1 and the non-corresponding model is represented by 0. In the example shown in FIG. 10, information indicating the drive system of the rotary machine 2 is included, and it is specified whether the drive system of the rotary machine 2 is a motor direct connection, a belt drive, an inverter drive, or the like. Information is included in the attribute information. For example, the attribute information includes information in which the corresponding drive system is represented by 1 and the non-corresponding drive system is represented by 0. Further, in the example shown in FIG. 10, the attribute information includes information for designating which type of bearing of the rotating machine is, such as a slide bearing or a rolling bearing. For example, the attribute information includes information in which the corresponding bearing type is represented by 1 and the non-corresponding type is represented by 0.

In the example shown in FIG. 10, information indicating whether the rigidity of the gantry / foundation on which the rotating machine 2 is installed is rigid or flexible is included. For example, the attribute information includes information in which the corresponding rigidity is represented by 1 and the non-corresponding rigidity is represented by 0. The attribute information may include other information as described above. Further, the attribute information may include information indicating the position where the vibration sensor 31 is attached. FIG. 10 shows an example in which both the information determined by the specifications and the information determined by the environment are included in the attribute information, but the attribute information may include only one of the information.

The attribute information is stored in the storage unit 13 in advance, and in S107, the calculation unit 11 reads the attribute information from the storage unit 13. Alternatively, in S107, the attribute information may be input by the operation unit 15 when the user operates the operation unit 15, and the calculation unit 11 may acquire the input attribute information.

Next, the calculation unit 11 uses the diagnostic model 133 to generate factor information regarding the factors of the vibration state (S108). The diagnostic model 133 is a learning model using NN. The diagnostic model 133 may use a CNN or a recurrent neural network (RNN) as the NN. The diagnostic model 133 is preferably an NN having a plurality of intermediate layers. The diagnostic model 133 may be a learning model other than the NN.

As shown in FIG. 7, the diagnostic model 133 is input with a plurality of state data and attribute information output by the vibration determination model 132, and outputs factor information. FIG. 11 is a conceptual diagram showing an example of the content of the factor information. The factor information includes the probability or certainty that the rotating machine 2 is normal. Further, in the factor information, regarding the vibration generated in the rotating machine 2, the position of the rotating shaft is deviated from the correct position, the bearing is not lubricated well, or the rotating machine 2 is mounted on the gantry. It includes the probability or certainty that it is caused by improper fixing that is not securely fixed. The diagnostic model 133 has been learned in advance so as to output factor information when a plurality of state data and attribute information are input.

In S108, the calculation unit 11 inputs a plurality of state data and attribute information output by the vibration determination model 132 into the diagnostic model 133, and acquires the factor information output by the diagnostic model 133. In this way, in S108, the calculation unit 11 generates factor information.

Next, the calculation unit 11 acquires a similar case from the case database 432 stored in the storage unit 43 of the learning device 4 (S109). In S109, the calculation unit 11 causes the communication unit 18 to transmit a request for a similar case to the learning device 4. For example, the request for similar cases includes information on diagnosis such as state data and factor information. The communication unit 18 transmits a request for a similar case to the learning device 4 via the communication network N. The learning device 4 receives a request for a similar case at the communication unit 45. The calculation unit 41 extracts case data including a diagnosis result having a degree of similarity with the received factor information within a predetermined range from the case database 432. For example, the calculation unit 41 narrows down the case data related to a model equivalent to the rotary machine 2, and determines the degree of similarity between the state data and factor information included in the narrowed down data and the state data and factor information included in the request for a similar case. calculate. For example, the difference between the vectors representing the state data and the factor information is calculated, and the absolute value of the vector difference is taken as the similarity. For example, the calculation unit 41 extracts data of a predetermined number of cases having the latest diagnosis date and time from the data of cases in which the similarity is equal to or less than a predetermined value. The calculation unit 41 causes the communication unit 45 to transmit the extracted data to the diagnostic device 1. By receiving the data from the learning device 4 in the communication unit 18, the calculation unit 11 acquires a similar case.

Next, the calculation unit 11 outputs a waveform representing a time change of the acceleration of vibration, factor information, and a similar case (S110). In S110, the calculation unit 11 causes the display unit 16 to display an image showing a waveform of vibration acceleration, factor information, and similar cases. FIG. 12 is a schematic view showing an example of an image displayed by the display unit 16. The image includes the waveform of the acceleration of the vibration represented by the time-series acceleration data 134, the frequency spectrum of the acceleration, and the content of analyzing the factors of the vibration. The vibration factor analysis results are displayed in the form of a bar graph with the class corresponding to each element of the factor information on the horizontal axis and the certainty on the vertical axis. For example, class 0 corresponds to the normality of the rotating machine 2, and class 1 corresponds to the misalignment. In the example shown in FIG. 12, the certainty of the class corresponding to "normal" is larger than that of the other classes and is almost the same value as 1, so that the state of the rotating machine 2 that causes vibration is "normal". The result of the analysis is displayed. When there are a plurality of classes having a high degree of certainty, the calculation unit 11 may display a plurality of vibration factors together on the display unit 16. The user can confirm the obtained factor information from the image displayed on the display unit 16.

Further, in the example shown in FIG. 12, the contents of similar cases in which the diagnosis result is similar to the factor information are displayed. By confirming the contents of the displayed similar cases, the user can confirm the past diagnosis results and can guess under what circumstances the acquired vibration is likely to occur. .. The calculation unit 11 may output a plurality of acquired similar cases.

Next, the calculation unit 11 waits for the acceptance of the correction of the factor information (S111). After checking the factor information included in the image displayed on the display unit 16, the user operates the operation unit 15 to correct the factor information when he / she feels that the obtained factor information is incorrect. You can enter it. When the correction of the factor information is accepted (S111: YES), the calculation unit 11 causes the communication unit 45 to transmit the data representing the content of the correction of the received factor information to the learning device 4 (S112). As will be described later, the learning device 4 learns the learning model based on the correction of the factor information. The calculation unit 11 ends the diagnosis process after S112 is completed.

After the end of S104 or when the correction of the factor information is not received in S111 (S111: NO), the calculation unit 11 outputs a diagnosis report including the result of diagnosing the rotary machine 2 (S113). In S113, the calculation unit 11 creates a diagnosis report according to the factor information, and causes the communication unit 18 to transmit the diagnosis report data to the output device 33. The output device 33 such as a printer receives the data of the diagnosis report from the diagnosis device 1 via the communication network N and outputs the diagnosis report. For example, the calculation unit 11 outputs a diagnostic report of the content that reports that the rotating machine 2 is normal because the amplitude of vibration is small according to the determination in S104. For example, the calculation unit 11 outputs a diagnostic report recording the same contents as the image output in S110 as shown in FIG. The calculation unit 11 may output a diagnostic report to which a sentence explaining the content of the factor information in words is added. The user can confirm the diagnosis result of the rotary machine 2 from the diagnosis report. In addition, the user can explain the diagnosis result of the rotary machine 2 to another vehicle by using the diagnosis report. The calculation unit 11 ends the diagnosis process after S113 is completed.

When a plurality of vibration sensors 31 are attached to the rotating machine 2, the calculation unit 11 executes the processes of S101 to S113 for each of the plurality of vibration sensors 31. The processes for the plurality of vibration sensors 31 may be sequentially performed or may be performed in parallel. In S110, the waveforms of vibration accelerations, factor information, and similar cases of the plurality of vibration sensors 31 may be collectively output.

Of S101 to S113, the process of S109 may be omitted. In this case, similar cases are not output in S110. Further, the processing of S113 may be performed after S112. Alternatively, the processing of S111 and S112 may be omitted.

The learning device 4 executes the learning model generation method. FIG. 13 is a conceptual diagram showing an example of the contents of the case database 432. The case database 432 records a large number of cases in which a plurality of rotating machines have been diagnosed in the past. Attribute information, time-series acceleration data, state data, and factor information related to the rotated machine that received the diagnosis are recorded in association with the date and time when the diagnosis was made. The attribute information includes information such as the model of the rotating machine, the driving method of the rotating machine, and the type of bearing. In the example shown in FIG. 13, the pump is recorded as a model. The case database 432 records time-series acceleration data acquired by a vibration sensor attached to the rotated machine that has been diagnosed. It should be noted that the first feature data such as the time series velocity data or the second feature data such as the frequency spectrum or the probability distribution of the value of the time series acceleration data may be further recorded.

Further, in the case database 432, state data representing the state of vibration is recorded. As shown in FIG. 13, a plurality of state data such as a state data representing a state of vibration acceleration and a state data representing a state of vibration velocity are recorded in the case database 432. Further, the case database 432 records factor information regarding the factors of the vibration state. These data are associated with each other and are recorded as examples of each of the previous diagnostics of the rotating machine. The contents of the state data and the personnel information recorded in the case database 432 are, for example, the contents determined by the person who observed the vibration using the vibration sensor in the past diagnosis.

The learning device 4 uses the data recorded in the case database 432 as training data to generate a vibration determination model 132 and a diagnostic model 133, which are learning models. FIG. 14 is a flowchart showing a procedure of processing in which the learning device 4 generates a set of NNs including a plurality of CNNs and one FCNN included in the vibration determination model 132. The calculation unit 41 reads out a plurality of time-series acceleration data and specific state data recorded for one model from the case database 432 stored in the storage unit 43 (S21). The specific state data is the kind of state data to be output by the set of NNs to be generated. For example, when generating a set of NNs including CNN511a, 511b and 511c and FCNN512 shown in FIG. 8, state data representing the state of vibration acceleration is read out.

Next, the calculation unit 41 generates necessary feature data from the plurality of read time-series acceleration data (S22). In S22, the calculation unit 41 generates the first feature data or the second feature data in the same manner as the method performed by the calculation unit 11 in S102 or S105. The feature data to be generated is the feature data to be input to the set of NNs to be generated. For example, when generating the NN set shown in FIG. 8, the frequency spectrum of the time series envelope data is generated. When the required feature data is recorded in the case database 432, the calculation unit 41 may perform a process of reading the required feature data from the case database 432 instead of the process of S22.

The calculation unit 41 uses a plurality of combinations of the read time-series acceleration data, the generated feature data, and the state data as training data, and outputs a set of NNs that output the time-series acceleration data and the state data when the feature data is input. Is generated (S23). In S23, the calculation unit 41 inputs the corresponding data among the time-series acceleration data, the first feature data, and the second feature data into each of the plurality of CNNs, and outputs a plurality of feature vectors output by the plurality of CNNs. Input to FCNN. For example, when generating the NN set shown in FIG. 8, the time series acceleration data is input to CNN511a, the frequency spectrum of acceleration is input to CNN511b, and the probability distribution of acceleration amplitude is input to CNN511c. The state data is output by FCNN. The calculation unit 41 calculates an error by an error function using the output state data and the state data used as training data as variables, and inputs a plurality of CNNs and FCNNs so that the error is minimized by the error back propagation method. Adjust the calculation parameters of each node included in the layer, intermediate layer and output layer.

The calculation unit 41 repeats the process using a plurality of combinations of the time-series acceleration data and the feature data and the state data included in the training data, and adjusts the parameters of the plurality of CNNs and FCNNs to form a set of NNs. Do machine learning. The calculation unit 41 generates a plurality of CNNs and FCNNs having adjusted final parameters.

After the end of S23, the calculation unit 41 ends the process. The calculation unit 41 executes the processes of S21 to S23 for each of the plurality of sets of NNs. By executing the processes of S21 to S23 a plurality of times, the calculation unit 41 generates a vibration determination model 132 including a plurality of sets of NNs including a plurality of CNNs and one FCNN.

FIG. 15 is a flowchart showing the procedure of the process in which the learning device 4 generates the diagnostic model 133. The calculation unit 41 reads out a plurality of state data, attribute information, and factor information recorded for one model from the case database 432 stored in the storage unit 43 (S31). The calculation unit 41 uses a plurality of combinations of the read state data, attribute information, and factor information as training data, and generates a diagnostic model 133 that outputs personnel information when the state data and attribute information are input (S32).

In S32, the calculation unit 41 inputs the state data and the attribute information to the diagnostic model 133. Factor information is output from the diagnostic model 133. The calculation unit 41 calculates the error by an error function using the output factor information and the factor information used as training data as variables, and the input layer of the diagnostic model 133 so that the error is minimized by the error back propagation method. Adjust the calculation parameters of each node included in the intermediate layer and output layer. The calculation unit 41 repeats the process using a plurality of combinations of state data, attribute information, and factor information included in the training data, and adjusts the parameters of the diagnostic model 133 to perform machine learning of the diagnostic model 133. The calculation unit 41 generates a diagnostic model 133 having the adjusted final parameters. After the end of S32, the calculation unit 41 ends the process.

The calculation unit 41 performs the processes of S21 to S23 and S31 to S32 for each model of the rotating machine to generate the vibration determination model 132 and the diagnostic model 133. That is, the learning device 4 reads the training data from the case database 432 for each model of the rotating machine and generates a learning model. The calculation unit 41 stores the data necessary for realizing the generated vibration determination model 132 and the diagnosis model 133 in the storage unit 43.

The calculation unit 41 causes the communication unit 45 to transmit the data necessary for realizing the generated vibration determination model 132 and the diagnosis model 133 to the diagnosis device 1. The diagnostic device 1 receives the data necessary for realizing the vibration determination model 132 and the diagnostic model 133 in the communication unit 18, and the calculation unit 11 stores the received data in the storage unit 13 to store the received data in the storage unit 13 to obtain the vibration determination model. 132 and diagnostic model 133 are realized. The storage unit 13 stores the data of the vibration determination model 132 and the diagnostic model 133 for each model of the rotary machine, and the diagnostic device 1 performs the diagnosis process using the learning model corresponding to the model of the rotary machine 2.

The learning device 4 can also relearn the learning model. The diagnostic device 1 transmits data representing the content of the correction of the factor information to the learning device 4 in the process of S112. The data transmitted by the diagnostic apparatus 1 in S112 includes attribute information, state information, and corrected factor information regarding the rotating machine 2. The learning device 4 receives the data in the communication unit 45, and the calculation unit 41 stores the received data in the correction database 433 stored in the storage unit 43. A plurality of combinations of attribute information, state information, and modified factor information are recorded in the modification database 433.

FIG. 16 is a flowchart showing the procedure of the re-learning process performed by the learning device 4. The calculation unit 41 reads training data composed of a plurality of combinations of attribute information, state information, and factor information from the correction database 433 (S41). The calculation unit 41 relearns the diagnostic model 133 based on the training data (S42). In S42, the calculation unit 41 inputs state data and attribute information to the diagnostic model 133, calculates an error by an error function using the factor information output from the diagnostic model 133 and the factor information used as training data as variables. The parameters of the diagnostic model 133 are adjusted so that the error is minimized by the error backpropagation method. The calculation unit 41 repeats the process using a plurality of combinations of the state data, the attribute information, and the factor information included in the training data, and relearns the diagnostic model 133. After the end of S42, the calculation unit 41 ends the process.

The calculation unit 41 stores the data necessary for realizing the relearned diagnostic model 133 in the storage unit 43. The calculation unit 41 causes the communication unit 45 to transmit the data necessary for realizing the relearned diagnostic model 133 to the diagnostic device 1. The diagnostic device 1 receives the data necessary for realizing the relearned diagnostic model 133 in the communication unit 18, and the arithmetic unit 11 updates the diagnostic model 133 by storing the received data in the storage unit 13. do. Subsequent diagnostic devices 1 perform diagnostic processing using the relearned diagnostic model 133. Re-learning of the diagnostic model 133 improves the accuracy of the diagnosis.

As described in detail above, the diagnostic apparatus 1 generates a plurality of types of feature data from the time-series acceleration data that records the acceleration of the vibration generated in the rotating machine 2. The feature data includes a first feature data that records an amount other than the acceleration representing vibration, or a second feature data relating to the frequency distribution or the probability distribution of the amplitude of the time-series acceleration data or the first feature data. The diagnostic apparatus 1 diagnoses the rotating machine 2 by using not only the time-series acceleration data but also the first feature data or the feature data including the second feature data. Similar to a human diagnosis using a plurality of types of information, the diagnostic device 1 can perform a diagnosis of the rotating machine 2 from a multimodal viewpoint using a plurality of types of information regarding vibration generated in the rotating machine 2. As a result, the diagnostic device 1 can perform the diagnosis more accurately than the conventional diagnosis. Further, the diagnostic device 1 diagnoses the rotating machine 2 based on the measurement result of vibration by using the data recorded in the case database 432 as training data without using the physical model of the rotating machine 2. be able to.

By using the vibration determination model 132, which is a learning model, state data representing the vibration state can be easily obtained from a plurality of types of information regarding vibration. The state data clarifies what kind of state the acceleration or velocity of vibration is in, such as a normal state or a state in which a specific frequency component is large. Further, by using the diagnostic model 133 which is a learning model, the factor information regarding the factor of the vibration state can be easily obtained from the state data and the attribute information regarding the rotating machine 2. From the factor information, it becomes clear what kind of state the rotating machine 2 is in as a factor in the vibration state. Therefore, it is possible to accurately diagnose what kind of state the rotating machine 2 is in.

The learning device 4 may be in a form of providing the vibration determination model 132 and the diagnostic model 133 that have been learned for the plurality of diagnostic devices 1. For example, a plurality of diagnostic devices 1 are connected to the communication network N, and the learning device 4 transmits data necessary for realizing the vibration determination model 132 and the diagnostic model 133 to the plurality of diagnostic devices 1. The learning device 4 may receive corrections of factor information from a plurality of diagnostic devices 1.

In the first embodiment, the diagnostic device 1 and the learning device 4 are connected to each other via the communication network N, but the diagnostic device 1 and the learning device 4 may be separated from each other. .. In this embodiment, data necessary for realizing the vibration determination model 132 and the diagnostic model 133 is passed from the learning device 4 to the diagnostic device 1 by a method other than communication via the communication network N. For example, the data necessary for realizing the vibration determination model 132 and the diagnostic model 133 are read from the learning device 4, recorded on a recording medium such as a portable memory or an optical disk, and read from the recording medium to the diagnostic device 1. ..

<Embodiment 2>
FIG. 17 is a schematic diagram showing the configuration of the diagnostic system according to the second embodiment. In the second embodiment, the A / D converter 32 is connected to the transmission unit 34. The transmission unit 34 is connected to the communication network N. The transmission unit 34 receives the signal output by the A / D converter 32, and transmits the received signal to the diagnostic device 1 via the communication network N. The diagnostic device 1 is a computer such as a server device. The diagnostic device 1 receives the signal transmitted from the transmission unit 34. That is, the diagnostic device 1 receives a signal indicating the value of the acceleration of vibration measured by the vibration sensor 31 via the communication network N. The diagnostic system includes a terminal device 6. The terminal device 6 is connected to the communication network N. The configuration of the other parts of the diagnostic system is the same as in the first embodiment.

The terminal device 6 is a device used by the user for diagnosing the rotating machine 2. For example, the terminal device 6 is arranged in the vicinity of the rotating machine 2. FIG. 18 is a block diagram showing an example of an internal functional configuration of the terminal device 6. The terminal device 6 is configured by using a computer such as a personal computer, a tablet computer, or a smartphone. The terminal device 6 includes a calculation unit 61, a memory 62, a storage unit 63, an operation unit 64, a display unit 65, and a communication unit 66. The calculation unit 61 is configured by using, for example, a CPU, a GPU, or a multi-core CPU. The memory 62 is, for example, a RAM. The storage unit 63 is non-volatile, for example, a hard disk. The operation unit 64 is, for example, a keyboard, a pointing device, or a touch panel. The display unit 65 is, for example, a liquid crystal display or an EL display. The communication unit 66 is connected to the communication network N and communicates with the outside of the terminal device 6. The communication unit 66 transmits / receives data to / from the diagnostic device 1 through the communication network N.

The storage unit 63 stores the computer program 631. For example, the computer program 631 is a web application program for diagnosing the rotating machine 2, is downloaded from the diagnostic device 1, and is stored in the storage unit 63. The computer program 631 may be stored in the memory 62.

The diagnostic device 1 receives a signal indicating the value of the acceleration of the vibration measured by the vibration sensor 31, and stores the time-series acceleration data 134 in the storage unit 13. The user operates the operation unit 64 to input a diagnosis instruction of the rotary machine 2. The calculation unit 61 causes the communication unit 66 to transmit an instruction to start diagnosis to the diagnosis device 1 in response to the input instruction. The diagnostic device 1 receives the instruction to start the diagnosis in the communication unit 18, and the calculation unit 41 executes the processes of S101 to S113. In S110, the calculation unit 11 causes the communication unit 18 to transmit a waveform representing the time change of the vibration acceleration, factor information, and a similar case to the terminal device 6, and in the terminal device 6, the waveform of the vibration acceleration, the factor information, And similar cases are displayed on the display unit 65. The user can operate the terminal device 6 to know the result of the diagnosis of the rotating machine 2.

The diagnostic device 1 may be in the form of diagnosing a plurality of rotating machines 2. For example, a plurality of transmission units 34 and a plurality of terminal devices 6 are connected to the communication network N. The diagnostic device 1 receives a signal indicating the value of the acceleration of the vibration measured by the plurality of vibration sensors 31, and stores the plurality of time-series acceleration data 134 in the storage unit 13. The diagnostic device 1 receives instructions for starting diagnosis from a plurality of terminal devices 6 and diagnoses each of the rotating machines 2. Further, the learning device 4 can receive corrections of factor information related to the plurality of rotating machines 2, record them in the correction database 433, and perform relearning processing.

Also in the second embodiment, the diagnostic device 1 can perform the diagnosis more accurately than the conventional diagnosis by using a plurality of types of information regarding the vibration generated in the rotating machine 2. In the second embodiment, the user does not need to directly operate the diagnostic device 1 and can more easily diagnose the rotary machine 2. Further, the learning device 4 can relearn the diagnostic model 133 based on the diagnostic results of the plurality of rotating machines 2. As a result, the accuracy of diagnosis using the diagnostic model 133 is further improved.

In the first and second embodiments, the form in which the vibration measurement result measured by the vibration sensor 31 is acquired by the diagnostic device 1 is shown, but the device for acquiring the vibration measurement result and the process for diagnosing the rotating machine 2 are performed. It may be a device different from the device to perform. In the first and second embodiments, the learning device 4 stores the case database 432, but the device that stores the case database 432 and the device that performs the learning process of the learning model may be different devices. ..

In the first and second embodiments, the vibration determination model 132 shows a form in which a plurality of sets of NNs including a plurality of CNNs and one FCNN are included, but the vibration determination model 132 includes a plurality of sets of CNNs. And may be in the form of only one set of NNs consisting of one FCNN. In the first and second embodiments, a mode in which the first feature data such as the time series velocity data is used in addition to the time series acceleration data 134 is shown, but the diagnostic apparatus 1 is a mode in which the first feature data is not used. May be good.

In the first and second embodiments, the first feature data such as the time-series velocity data is calculated from the time-series acceleration data 134, but the diagnostic apparatus 1 is a method other than the method of calculating from the time-series acceleration data 134. It may be in the form of acquiring the first feature data in. For example, a sensor for measuring the vibration speed of the rotating machine 2 may be attached to the rotating machine 2, and the diagnostic device 1 may acquire time-series speed data in which the speed measured by the sensor is recorded in time series. The diagnostic device 1 may be in the form of using the first feature data other than the time series velocity data, the time series envelope data, and the time series displacement data. In addition to the frequency spectrum and the probability distribution of amplitude, the diagnostic device 1 may be in the form of using feature data showing the characteristics of vibration.

The present invention is not limited to the contents of the above-described embodiment, and various modifications can be made within the scope of the claims. That is, an embodiment obtained by combining technical means appropriately modified within the scope of the claims is also included in the technical scope of the present invention.

1 Diagnostic device 131 Computer program 132 Vibration judgment model 133 Diagnostic model 2 Rotating machine 31 Vibration sensor 33 Output device 4 Learning device 431 Computer program 6 Terminal device N Communication network

The present invention relates to a processing method for diagnosing the state of a rotating machine from the measured vibration of the rotating machine, a method for diagnosing the rotating machine, a computer program, a learning model generation method, and a diagnostic device.

The present invention has been made in view of such circumstances, and an object of the present invention is a processing method capable of accurately determining the type of abnormality of a rotating machine, a diagnostic method of the rotating machine, a computer program, and the like. The purpose of the present invention is to provide a learning model generation method and a diagnostic device.

Multi-processing method according to the present invention, which acquires a series acceleration data when the time series records the acceleration of vibrations is a vibration sensor mounted on the rotating turning point was measured and calculated in a different way from the time series acceleration data To the first learning model that generates the type of feature data and outputs the state data representing the state of the vibration when a plurality of types of data among the time series acceleration data and the plurality of types of feature data are input. By inputting the plurality of types of data, the first training model includes a plurality of sets including a plurality of convolutional neural networks and one fully connected neural network, and a plurality of convolutions included in each set. The neural network has a one-to-one correspondence with a plurality of specific types of data among the plurality of types of data, and outputs a feature vector of the plurality of specific types of data when the corresponding specific types of data are input. However, the fully connected neural network included in each set outputs the state data when a plurality of feature vectors output by a plurality of convolutional neural networks included in the same set are input, and is included in each set. It is characterized in that the state data output by the fully connected neural network is acquired.
The method for diagnosing a rotating machine according to the present invention is a method for diagnosing a rotating machine by acquiring time-series acceleration data in which the acceleration of vibration measured by a vibration sensor attached to the rotating machine is recorded in time series. When a plurality of types of feature data calculated by different methods are generated from the time series acceleration data and a plurality of types of data among the time series acceleration data and the plurality of types of feature data are input, the vibration state is shown. The plurality of types of data are input to the first learning model that outputs the obtained state data, the state data output by the first learning model is acquired, the attribute information of the rotating machine is acquired, and the state data and the state data and the state data are acquired. The state data and the attribute information are input to the second learning model that outputs the factor information related to the factor of the vibration state when the attribute information is input, and the factor information output by the second learning model is acquired. It is characterized by doing.

Processing method of the rotating machine according to the present invention obtains the series acceleration data when the time series records the acceleration of vibrations is a vibration sensor mounted on the rotating turning point was measured, the amount representing the vibration of the non-acceleration The first feature data recorded in time series is acquired, the second feature data is generated from the time series acceleration data or the first feature data, and the time series acceleration data, the first feature data and the second feature data are generated. If you enter a plurality of types of data of the characteristic data, to the first learning model for outputting status data representing the state of the vibration type the plurality of types of data, the first training model, multiple The convolutional neural network includes a plurality of sets including one fully connected neural network, and the plurality of convolutional neural networks included in each set includes a plurality of specific types of the plurality of types of data. There is a one-to-one correspondence with the data, and when a specific type of data corresponding to each is input, the feature vectors of the plurality of specific types of data are output, and the fully connected neural networks included in each set are the same set. When a plurality of feature vectors output by a plurality of convolutional neural networks included in the data are input, the state data is output, and the state data output by the fully connected neural network included in each set is acquired. do.
The method for diagnosing a rotating machine according to the present invention is a method for diagnosing a rotating machine by acquiring time-series acceleration data in which the acceleration of vibration measured by a vibration sensor attached to the rotating machine is recorded in time series. The first feature data in which the amount representing the vibration other than the above is recorded in time series is acquired, the second feature data is generated from the time series acceleration data or the first feature data, and the time series acceleration data, the said. When a plurality of types of data among the first feature data and the second feature data are input, the plurality of types of data are input to the first learning model that outputs the state data representing the vibration state. The state data output by the first learning model is acquired, the attribute information of the rotating machine is acquired, and when the state data and the attribute information are input, the factor information regarding the factor of the vibration state is output. 2. The state data and the attribute information are input to the learning model, and the factor information output by the second learning model is acquired.

The computer program according to the present invention acquires time-series acceleration data in which the acceleration of vibration measured by a vibration sensor attached to a rotating machine is recorded in time series, and is calculated by a different method from the time-series acceleration data. To the first learning model that generates the type of feature data and outputs the state data representing the state of the vibration when a plurality of types of data among the time series acceleration data and the plurality of types of feature data are input. A computer is made to execute a process of inputting the plurality of types of data and acquiring the state data output by the first learning model, and the first learning model includes a plurality of convolutional neural networks and one fully connected neural network. A plurality of convolutional neural networks included in each set include a plurality of sets of data, and each of the plurality of convolutional neural networks has a one-to-one correspondence with a plurality of specific types of data among the plurality of types of data. When a specific type of data is input, the feature vectors of the plurality of specific types of data are output, and the fully connected neural network included in each set is a plurality of output by a plurality of convolutional neural networks included in the same set. It is characterized in that the state data is output when the feature vector of is input .
The computer program according to the present invention acquires time-series acceleration data in which the acceleration of vibration measured by a vibration sensor attached to a rotating machine is recorded in time series, and is calculated by a different method from the time-series acceleration data. To the first learning model that generates the type of feature data and outputs the state data representing the state of the vibration when a plurality of types of data among the time series acceleration data and the plurality of types of feature data are input. Enter the multiple types of data and
The state data output by the first learning model is acquired, the attribute information of the rotating machine is acquired, and when the state data and the attribute information are input, the factor information regarding the factor of the vibration state is output. 2. It is characterized in that a process of inputting the state data and the attribute information into the learning model and acquiring the factor information output by the second learning model is executed by a computer.

In the computer program according to the present invention, wherein data of said plurality of types that have been generated, characterized in that it comprises a first feature data representing the time variation of the quantity representing the vibration other than acceleration.

A computer program according to the present invention, from the database recording the plurality of cases including a prior diagnosis result, the diagnostic result as the similarity of the rotating machine in accordance with the contents of the factor information obtains the case in a predetermined range It is characterized in that the computer further executes the process of outputting the acquired case.

The learning model generation method according to the present invention is calculated by a method different from the time-series acceleration data in which the acceleration of vibration measured in the past by the vibration sensor attached to the rotating machine is recorded in time series and the time-series acceleration data. Tokushi preparative plurality of types of characteristic data for the time series acceleration data and the plurality of types of characteristic data, obtains the state data representing the state of the vibration, the time series acceleration data and the plurality of types of feature When the time-series acceleration data in which the acceleration of an arbitrary vibration is recorded in time series and a plurality of types of feature data generated from the time-series acceleration data are input by using the data and the state data as training data, the said A first learning model that outputs state data representing an arbitrary vibration state is generated , and attribute information of the rotating machine and factor information related to factors of the vibration state measured in the past by the vibration sensor are acquired. , The state data, the attribute information, and the factor information are used as training data, and the state data representing the state of vibration measured by the vibration sensor attached to the arbitrary rotating machine and the attribute information of the arbitrary rotating machine are input. It is characterized by generating a second learning model that outputs factor information regarding the factors of the vibration state in the case.

The learning model generation method according to the present invention obtains the time-series acceleration data by reading the time-series acceleration data from a database recording a plurality of cases including past diagnosis results without using a physical model. It is characterized in that the state data , the attribute information and the factor information are acquired by acquiring and reading the state data , the attribute information and the factor information from the database.

The diagnostic device according to the present invention is a diagnostic device for diagnosing a rotating machine, and includes an acceleration acquisition unit that acquires time-series acceleration data in which the acceleration of vibration measured by a vibration sensor attached to the rotating machine is recorded in time series. The vibration when a generator that generates a plurality of types of feature data calculated by different methods from the time-series acceleration data and a plurality of types of data among the time-series acceleration data and the plurality of types of feature data are input. first learning model and the plurality of types of data input to the first learning model and a state data acquisition unit for acquiring the status data, the first learning model for outputting status data representing the status Includes a plurality of sets including a plurality of convolutional neural networks and one fully connected neural network, and the plurality of convolutional neural networks included in each set includes a plurality of the plurality of types of data. There is a one-to-one correspondence with specific types of data, and when the corresponding specific types of data are input, the feature vectors of the plurality of specific types of data are output, and the fully connected neural network included in each set is It is characterized in that the state data is output when a plurality of feature vectors output by a plurality of convolutional neural networks included in the same set are input.
The diagnostic device according to the present invention is a diagnostic device for diagnosing a rotating machine, and includes an acceleration acquisition unit that acquires time-series acceleration data in which the acceleration of vibration measured by a vibration sensor attached to the rotating machine is recorded in time series. The vibration when a generator that generates a plurality of types of feature data calculated by different methods from the time-series acceleration data and a plurality of types of data among the time-series acceleration data and the plurality of types of feature data are input. A first learning model that outputs state data representing the state of the above, a state data acquisition unit that inputs the plurality of types of data into the first learning model to acquire the state data, and attribute information of the rotating machine. The attribute information acquisition unit to be acquired, the second learning model that outputs the factor information related to the factor of the vibration state when the state data and the attribute information are input, and the second learning of the state data and the attribute information. It is characterized by including a factor information acquisition unit that inputs to a model and acquires the factor information.

Claims (19)

In the method of diagnosing a rotating machine,
Acquire time-series acceleration data that records the acceleration of vibration measured by the vibration sensor attached to the rotating machine in chronological order.
Multiple types of feature data are generated from the time-series acceleration data,
When a plurality of types of data among the time-series acceleration data and the plurality of types of feature data are input, the plurality of types of data are input to the first learning model that outputs the state data representing the state of the vibration. ,
A method for diagnosing a rotating machine, which comprises acquiring the state data output by the first learning model. In the method of diagnosing a rotating machine,
Acquire time-series acceleration data that records the vibration acceleration measured by the vibration sensor attached to the rotating machine in chronological order.
Acquire the first feature data in which quantities other than the acceleration representing the vibration are recorded in time series, and
The second feature data is generated from the time series acceleration data or the first feature data.
When a plurality of types of data among the time-series acceleration data, the first feature data, and the second feature data are input, the plurality of data are added to a first learning model that outputs state data representing the state of vibration. Enter the type of data,
A method for diagnosing a rotating machine, which comprises acquiring the state data output by the first learning model. Acquire time-series acceleration data that records the vibration acceleration measured by the vibration sensor attached to the rotating machine in chronological order.
Multiple types of feature data are generated from the time-series acceleration data,
When a plurality of types of data among the time-series acceleration data and the plurality of types of feature data are input, the plurality of types of data are input to the first learning model that outputs the state data representing the state of the vibration. ,
A computer program characterized in that a computer executes a process of acquiring the state data output by the first learning model. The computer program according to claim 3, wherein the plurality of types of feature data include first feature data in which quantities other than acceleration representing the vibration are recorded in time series. The first feature data is time-series velocity data in which the velocity of the vibration is recorded in time series, time-series envelope data in which the envelope of the acceleration of the vibration is recorded in time series, or displacement of the vibration is timed. It is time series displacement data recorded in series,
A process of generating the time-series velocity data by integrating the time-series acceleration data,
The computer is subjected to a process of generating the time-series envelope data by generating the envelope from the time-series acceleration data, or a process of generating the time-series displacement data by integrating the time-series acceleration data twice. The computer program according to claim 4, wherein the computer program is executed. When the value of the time-series acceleration data is included in the predetermined first range, the value of the time-series velocity data is included in the predetermined second range, or the value of the time-series displacement data is predetermined. The computer program according to claim 5, wherein the computer further executes a process of determining that the rotating machine is normal when it is included in the third range of the above. The plurality of types of feature data are frequency spectra obtained from the time-series acceleration data or the first feature data, or second feature data which is a probability distribution of amplitude obtained from the time-series acceleration data or the first feature data. The computer program according to any one of claims 4 to 6, wherein the computer program comprises. The third aspect of claim 3, wherein the plurality of types of feature data include a frequency spectrum obtained from the time-series acceleration data or a second feature data which is a probability distribution of an amplitude obtained from the time-series acceleration data. Computer program. The first learning model has a one-to-one correspondence with the plurality of types of data, and a plurality of convolutional neural networks that output a feature vector of the data when the corresponding data is input, and the plurality of convolutional neural networks. The computer program according to any one of claims 3 to 8, wherein the computer program includes a fully connected neural network that outputs the state data when a plurality of feature vectors output by the user are input. The first learning model includes a plurality of sets including a plurality of convolutional neural networks and one fully connected neural network.
The plurality of convolutional neural networks included in each set have a one-to-one correspondence with a plurality of specific types of data among the plurality of types of data, and when the corresponding specific types of data are input, the plurality of convolutional neural networks are described. Outputs the feature vector of a specific type of data
Claims 3 to 3 to the fully coupled neural network included in each set output the state data when a plurality of feature vectors output by a plurality of convolutional neural networks included in the same set are input. The computer program according to any one of 8. Acquire the attribute information of the rotating machine and
When the state data and the attribute information are input, the state data and the attribute information are input to the second learning model that outputs the factor information regarding the factor of the vibration state.
The computer program according to any one of claims 3 to 10, wherein a computer further executes a process of acquiring the factor information output by the second learning model. The computer program according to claim 11, wherein a computer further executes a process of outputting a diagnostic report of the rotating machine according to the content of the factor information. Output the contents of the factor information
Accepting corrections to the factor information
The computer program according to claim 11 or 12, wherein the computer further executes a process of retraining the second learning model based on the received modification. From a database that records a plurality of cases including past diagnosis results, cases in which the degree of similarity with the diagnosis results for the rotary machine is within a predetermined range are acquired.
The computer program according to any one of claims 11 to 13, wherein a computer further executes a process of outputting the acquired case. Acquire multiple types of time-series acceleration data, which records the acceleration of vibration measured in the past by a vibration sensor attached to the rotating machine, and multiple types of feature data generated from the time-series acceleration data. ,
With respect to the time-series acceleration data and the plurality of types of feature data, state data representing the state of vibration is acquired.
Using the time-series acceleration data, the plurality of types of feature data, and the state data as training data, the time-series acceleration data in which the acceleration of an arbitrary vibration is recorded in a time-series manner, and a plurality of data generated from the time-series acceleration data. A learning model generation method, characterized in that, when inputting a type of feature data, a first learning model that outputs state data representing the state of the arbitrary vibration is generated. The time-series acceleration data is acquired by reading the time-series acceleration data from a database that records a plurality of cases including past diagnosis results without using a physical model.
The learning model generation method according to claim 15, wherein the state data is acquired by reading the state data from the database. A plurality of types of state data representing the vibration state measured in the past by the vibration sensor attached to the rotating machine, the attribute information of the rotating machine, and the factor information related to the factor of the vibration state are acquired.
When a plurality of types of the state data, the attribute information, and the factor information are used as training data, and the state data representing the vibration state of an arbitrary rotating machine and the attribute information of the arbitrary rotating machine are input, the vibration state. A learning model generation method characterized by generating a second learning model that outputs factor information related to the factors of. It is characterized in that the state data, the attribute information and the factor information are acquired by reading the state data, the attribute information and the factor information from a database recording a plurality of cases including past diagnosis results. The learning model generation method according to claim 17. In a diagnostic device that diagnoses rotating machines
An acceleration acquisition unit that acquires time-series acceleration data that records the vibration acceleration measured by a vibration sensor attached to the rotating machine in time series, and a generation unit that generates multiple types of feature data from the time-series acceleration data. ,
A first learning model that outputs state data representing the state of vibration when a plurality of types of data among the time-series acceleration data and the plurality of types of feature data are input.
A diagnostic apparatus including a state data acquisition unit that inputs the plurality of types of data and acquires the state data.

Images