JP7552366B2 - Abnormality diagnosis device, abnormality diagnosis system, and abnormality diagnosis method - Google Patents

Description

The present invention relates to a technology for diagnosing abnormalities in a diagnostic target such as a rotating machine, and in particular to a technology for selecting learning data to improve diagnostic accuracy in diagnosing abnormalities in a diagnostic target based on a constant Q transformation.

In recent years, methods have been proposed for diagnosing abnormalities in equipment such as rotating machines based on the operating state of the equipment during operation (for example, Patent Documents 1 to 5). In particular, the abnormality diagnosis in Patent Documents 1 to 3 is performed by a diagnosis based on a constant Q conversion of data from an acceleration sensor equipped in the rotating machine. As the basic diagnostic method is to learn only normal data, since there is extremely little abnormal data when the rotating machine continues to operate normally, and to diagnose whether the equipment is normal or abnormal by quantifying how much the current data differs from the normal data.

JP 2017-198619 A JP 2017-198620 A JP 2019-101777 A JP 2019-079356 A JP 2019-027860 A

However, the accuracy of conventional methods for diagnosing abnormalities in equipment is reduced due to the following reasons:

Since the normal state in which the equipment is operating stably is learned, unstable acceleration when the equipment is stopped (hereafter referred to as the "transition period") may be judged as abnormal.

Immediately after the acceleration sensor is installed in the device or after repairs are made to the device, the sensor output may not be stable for the following reasons:

The sensor is attached to the equipment using magnets, adhesives, bolts, etc., but the sensor output is not stable until the installation is stable.

When repairs are made, the rotating machine body is bolted to the base, so the sensor output will not stabilize until the bolts are tightened properly.

In addition, data from periods when the sensor output is unstable will show greater or smaller acceleration than during stable operation, so learning data from those periods will result in reduced accuracy.

Furthermore, as shown in Figure 32, when the operating data of a device for a certain period (one week or one month) is learned, annual fluctuations occur based on that period. This annual fluctuation can cause the threshold to be exceeded even when the device is normal, resulting in a false detection of an abnormality. In addition, annual fluctuations are thought to be caused by the expansion and contraction of devices due to temperature, and in the case of pumps, for example, the change in the amount of air dissolved in water.

In view of the above, the present invention aims to improve the accuracy of abnormality diagnosis based on constant Q transformation of the diagnosis target.

Therefore, one aspect of the present invention is an abnormality diagnosis device that sequentially learns the operating state of an object to be diagnosed and diagnoses abnormalities, and includes a transient period removal unit that removes acceleration data determined to be in a transient period of the operating state of the object to be diagnosed based on fluctuations in acceleration amplitude obtained from the acceleration data of the object to be diagnosed from the sequential learning.

One aspect of the present invention is an abnormality diagnosis device that sequentially learns the operating state of a diagnosis target and diagnoses abnormalities, and includes a data filtering unit that excludes acceleration data determined to be in a transitional period of the operating state of the diagnosis target based on peak fluctuations in a rotational frequency band obtained from the acceleration data of the diagnosis target from the sequential learning.

In one aspect of the present invention, the abnormality diagnosis device includes a learning period selection unit that selects a time period during which the standard deviation of the moving average of the acceleration effective value of the acceleration data converges to a predetermined range as the period of the sequential learning.

In one aspect of the present invention, the abnormality diagnosis device includes a sequential learning processing unit that starts the sequential learning in response to a detection signal of the operating state of another object related to the diagnosis object or a manual trigger signal that commands the sequential learning.

In one aspect of the present invention, the abnormality diagnosis device includes a sequential learning processing unit that learns the normal transition of the effective acceleration value when the operation of the diagnosis target is stopped.

In one aspect of the present invention, in the abnormality diagnosis device, the diagnosis target is a variable-speed rotating machine, and a specific rotation frequency set based on acceleration data of the rotating machine is used for the sequential learning and the diagnosis.

One aspect of the present invention is an abnormality diagnosis method that sequentially learns the operating state of an object to be diagnosed and diagnoses abnormalities, and excludes acceleration data that is determined to be in a transitional period of the operating state of the object to be diagnosed based on the fluctuation in acceleration amplitude or the peak fluctuation in the rotational frequency band obtained from the acceleration data of the object to be diagnosed from the sequential learning.

One aspect of the present invention is an abnormality diagnosis system that sequentially learns the operating state of a diagnosis target and diagnoses abnormalities, and includes the abnormality diagnosis device that performs the sequential learning and diagnosis, and a learning device that re-learns the criteria for the diagnosis based on the multivariate samples of the acceleration data obtained by the sequential learning and existing multivariate samples of the acceleration data.

In one aspect of the present invention, in the abnormality diagnosis system, the learning device performs the re-learning without performing a constant Q conversion of the time series data of the acceleration data.

In one aspect of the present invention, in the abnormality diagnosis system, the learning device performs the re-learning based on multivariate samples of the same or adjacent levels of the existing acceleration data obtained by constant Q transformation of the time series data of the acceleration data, and multivariate samples of the same or adjacent levels of the newly measured acceleration data obtained by constant Q transformation of the time series data of the acceleration data.

One aspect of the present invention is an abnormality diagnosis method for sequentially learning the operating state of a diagnosis target and diagnosing abnormalities, and includes a step of excluding acceleration data determined to be in a transitional period of the operating state of the diagnosis target based on the fluctuation in acceleration amplitude or the peak fluctuation in the rotational frequency band obtained from the acceleration data of the diagnosis target from the sequential learning.

In one aspect of the present invention, the abnormality diagnosis method includes a step of re-learning the diagnostic criteria based on the multivariate samples of the acceleration data obtained by the sequential learning and existing multivariate samples of the acceleration data.

According to the present invention, the accuracy of abnormality diagnosis based on constant Q transformation of the diagnosis target can be improved.

FIG. 1 is a configuration diagram of an equipment anomaly diagnosis system according to a first embodiment of the present invention. FIG. 1 is a configuration diagram of an equipment anomaly diagnosis system according to a first to third embodiment of the present invention. FIG. 1 is a block diagram of an abnormality diagnosis device according to a first embodiment. FIG. 13 is a block diagram of an abnormality diagnosis device according to the first to third embodiments. 4 is a flowchart of equipment abnormality diagnosis according to the first embodiment. Changes in acceleration data over time during the transient period of operation of the diagnostic target. The change over time in acceleration data in which the transient period in FIG. 6 is divided into multiple sections. FIG. 7 is a waveform diagram of a rotation frequency obtained by frequency analysis of a predetermined section in FIG. 6 . FIG. 9 is an explanatory diagram of determining a transitional period based on a comparison of the rotational frequency waveforms in a predetermined section of FIG. 8 . Change in effective value of acceleration data over time. Change over time in the calculated value of the moving average of the effective values in FIG. FIG. 12 is an explanatory diagram of a method for calculating the slope of the moving average in FIG. 11 . 13 is a flowchart of equipment abnormality diagnosis in Example 1-2. 11 is a flowchart of equipment abnormality diagnosis in the embodiment 1-3. FIG. 11 is an explanatory diagram of remeasurement and rediagnosis of acceleration data in Example 1-3. 13 is a flowchart of equipment abnormality diagnosis in the embodiment 1-4. Example of effective acceleration value when operation is stopped. An example of the change in deviation when operations are stopped. FIG. 5 is a configuration diagram of an equipment anomaly diagnosis system according to a second embodiment of the present invention. FIG. 11 is a block diagram of an abnormality diagnosis device according to a second embodiment. 13 is a flowchart of equipment abnormality diagnosis in Example 2-1. FIG. 13 is a block diagram of an abnormality diagnosis device according to an embodiment 2-4. 13 is a flowchart of an aperture parameter update process. This is a distribution diagram of the rotational frequency ratio for the entire period and the first month, estimated from data measured over a period of just over two years at a certain facility. This is a graph showing the results of a diagnosis without limiting the rotational frequency based on data measured over a period of just over two years for the equipment in Figure 24. A graph showing the results of a diagnosis with a rotation frequency limit of 19 Hz based on data collected over a period of just over two years on a certain piece of equipment. This is the result of frequency analysis when the rotational frequency of a variable speed rotating machine is 16 Hz and 20 Hz. FIG. 11 is a configuration diagram of an equipment anomaly diagnosis system according to a third embodiment of the present invention. FIG. 11 is a block diagram of an abnormality diagnosis device according to a third embodiment. 13 is a flowchart of equipment abnormality diagnosis according to the third embodiment. 13 is an explanatory diagram of data classification for each rotation frequency based on frequency analysis of measurement data in the third embodiment. Changes in deviation over time showing fluctuations in equipment operation data. FIG. 11 is a configuration diagram of an equipment anomaly diagnosis system according to a fourth embodiment of the present invention. FIG. 13 is a block diagram of an abnormality diagnosis device according to a fourth embodiment. 13 is a flowchart of equipment abnormality diagnosis according to the fourth embodiment. 13 is a flowchart of a sequential learning process according to the fourth embodiment. FIG. 13 is a configuration diagram of an equipment anomaly diagnosis system according to a fifth embodiment of the present invention. FIG. 13 is a block diagram showing the configuration of a learning device according to a fifth embodiment.

The following describes an embodiment of the present invention with reference to the drawings.

[Embodiment 1]
An equipment abnormality diagnosis system 2 to which an abnormality diagnosis device 1 according to one embodiment of the present invention shown in FIG. 1 is applied performs the following processes (1) to (3) when diagnosing an abnormality in a rotating machine 3 based on constant Q transformation (Patent Documents 1 to 3), thereby improving the accuracy of diagnosis of the rotating machine 3.

(1) By removing data on the transitional period of the operating state from the operation data of the rotating machine 3, erroneous detection of an abnormality in the rotating machine 3 due to the transitional period is prevented.

(2) Select the optimal period for learning the normal operating state of the rotating machine 3 by detecting the stabilization of the output value of the acceleration sensor 4 provided in the rotating machine 3.

(3) By sequentially learning the operating data of the rotating machine 3 and suppressing annual fluctuations, erroneous detection of abnormalities in the rotating machine 3 due to annual fluctuations is prevented.

The abnormality diagnosis device 1 is used to diagnose a rotating machine 3. The rotating machine 3 is connected to a rotating equipment 5 such as a pump, turbo blower, water wheel, etc. An acceleration sensor 4 is attached to the rotating machine 3. Acceleration data of the rotating machine 3 detected by the acceleration sensor 4 is output to the abnormality diagnosis device 1 via a data measurement unit 6. The data measurement unit 6 receives the output signal of the acceleration sensor 4 as acceleration data of the rotating machine 3 and provides it to the abnormality diagnosis device 1.

The abnormality diagnosis device 1 implements the following functional units shown in Figure 3 through cooperation between the computer's hardware and software resources.

The memory unit 10 stores the acceleration data provided by the data measurement unit 6.

The transient period removal unit 11 determines whether the operating state of the rotating machine 3 is in a transient period based on the acceleration data retrieved from the memory unit 10, and if it is in a transient period, excludes the acceleration data from the sequential learning. In particular, the acceleration data that is determined to be in a transient period of the operating state of the diagnosis target based on the fluctuation of the acceleration amplitude obtained from the acceleration data is excluded from the sequential learning.

The learning period selection unit 12 selects a period during which the output signal from the acceleration sensor 4 is stable as the learning period based on the acceleration data from which the acceleration data in the transient period has been removed. In particular, it selects a time period during which the standard deviation of the moving average of the effective acceleration value obtained from the acceleration data converges within a predetermined range as the period for the sequential learning.

The sequential learning processing unit 13 sequentially learns the normal operating state of the rotating machine 3 from the selected learning period. The timing of the learning depends on the performance of the abnormality diagnosis device 1, so sequential learning may be performed by adding one piece of data per day, or one piece of data per week. In addition, sequential learning must be performed for at least one year to suppress annual fluctuations, and after one year has passed, sequential learning may be continued or terminated.

The diagnostic processing unit 14 uses the learned learning result model to diagnose the state of the rotating machine 3 from the current acceleration data. In addition, if the diagnosis result is judged to be abnormal, measurement and diagnosis are performed again, and if it is judged to be abnormal continuously, it is judged to be truly abnormal, thereby improving the diagnostic accuracy. Basically, when an abnormality occurs in the equipment (rotating machine 3), the abnormality will continue to occur, so if it is judged to be abnormal continuously, it is judged to be abnormal.

The diagnosis result transmission unit 15 transmits the diagnosis results to a monitoring and control system, a cloud, etc.

(Example 1-1)
The diagnosis process of the first embodiment will be described with reference to the flowchart of FIG.

S111: An acceleration sensor 4 is installed, for example on a bearing, of the rotating machine 3 to be measured, and acceleration data is measured. The output signal of the acceleration sensor 4 is stored in the memory unit 10 as acceleration data of the rotating machine 3 via the data measurement unit 6.

S112: The transitional period removal unit 11 judges whether the acceleration data retrieved from the memory unit 10 is in a transitional period. Methods for judging whether the acceleration data is in a transitional period are shown in Figs. 6 to 9. When there is a transitional acceleration data as shown in Fig. 6, first, the acceleration data is divided every second as shown in Fig. 7. As shown in Fig. 8, a frequency analysis is performed for each divided acceleration data, and a rotation frequency is obtained for each divided acceleration data as shown in Fig. 9. If the obtained rotation frequency does not match, it is judged to be in a transitional period. No learning or diagnosis is performed on acceleration data judged to be in a transitional period. In addition, if the timing of the stop of operation of the equipment is known by linking with a monitoring and control system, data from the monitoring and control system may be used to judge whether the acceleration data is in a transitional period. If the period corresponding to the acceleration data is judged to be in a transitional period, it is removed from the learning period.

S113: If the transitional period has been removed in S112 (Yes), the process returns to S111. On the other hand, if the transitional period has not been removed (No), the process proceeds to the learning period selection process in S114.

S114: If it is determined that the period is not a transitional period, the learning period selection unit 12 performs a process for selecting a learning period. Basically, the learning period is determined to begin when the sensor output has stabilized. A method for determining whether the sensor output has stabilized is shown in Figures 10 to 12. Figure 10 shows an example of calculating the effective value of acceleration data, and the effective value becomes higher after the sensor is installed or repaired. Therefore, the moving average of the effective value of the acceleration data is calculated as shown in Figure 11, and the slope of the moving average is calculated as shown in Figure 12. The standard deviation of the slope is calculated, and the point at which the slope converges to within the range of "±σ" for one week is selected as the start of the learning period.

S115: If a learning period was not selected in S114 (No), the process returns to S111. On the other hand, if the learning period was selected (Yes), one-year sequential learning is performed in S116.

S116: If a learning period is selected in S115, the sequential learning processing unit 13 performs sequential learning every day for one year. For the learning, the method described in Patent Documents 1-3, for example, is used. Once one year's worth of sequential learning has been performed, the process proceeds to diagnosis processing in S118.

S117: If one year's worth of learning has not been performed, sequential learning is performed by the sequential learning processing unit 13. If one year's worth of learning has been performed, the diagnosis process of S118 is performed. Depending on the performance of the CPU of the abnormality diagnosis device 1 that performs the sequential learning, one piece of data may be learned per day, or one piece of data may be learned per week. After one year of learning, the influence of annual variation disappears, so sequential learning is no longer necessary, but may be continued.

S118: The diagnostic processing unit 14 performs diagnostic processing to determine whether the current acceleration data is normal or abnormal using the model learned in S117. For the diagnostic processing, the methods described in Patent Documents 1-3 are used.

S119: The diagnosis result transmission unit 15 executes a process for transmitting the diagnosis result obtained in S118 to a monitoring control system, a cloud, etc.

As described above, in the process of diagnosing abnormalities in equipment, vibrations that are different from normal during the transitional period are automatically removed, and an appropriate learning period is selected by automatically determining the period during which the rotating machine 3 is operating stably from the acceleration data. By performing sequential learning, the effects of annual fluctuations are eliminated, improving the accuracy of diagnosis.

When judging the transitional period, abnormal data and single abnormalities may be excluded. The measurement time for the acceleration data may be set according to the frequency analysis accuracy. Furthermore, the measurement period for detecting fluctuations in the rotation frequency may be set according to the system characteristics. Furthermore, when calculating the average of the frequency characteristics, weighting according to the measurement time may be applied. By applying these methods, overdetection during diagnosis can be more effectively avoided (the same applies to the second and third embodiments described below).

(Example 1-2)
In this embodiment, the accuracy of diagnosis is improved by excluding abnormal data from the targets of sequential learning.

The diagnosis process of Example 1-2 will be explained based on the flowchart in Figure 13. The process up to the learning period selection process (S111 to S115) is the same as in Example 1, so the explanation will be omitted. Below, the process after the learning period selection process (S116 to S125) will be explained.

S116, S117: The sequential learning processing unit 13 performs the first sequential learning as usual. Since the first sequential learning is generally performed after a sensor is installed or repaired, the device is treated as being in a normal operating state.

S118: The first diagnostic process is performed by the diagnostic processing unit 14.

S119: The first diagnosis result is transmitted by the diagnosis result transmission unit 15 to a monitoring control system, cloud, etc.

S121: From the second time onwards, the diagnostic processing unit 14 first performs diagnostic processing.

S122, S125: If it is determined that there is an abnormality (Yes), sequential learning is not performed, and the diagnosis result is transmitted by the diagnosis result transmission unit 15 to a monitoring control system, cloud, etc.

S122 to S124, S125: If it is determined that there is no abnormality (No), sequential learning is performed by the sequential learning processing unit 13. After that, the diagnosis result obtained by the diagnosis processing unit 14 is transmitted by the diagnosis result transmission unit 15 to a monitoring control system, a cloud, etc.

As described above, by automatically excluding abnormal data from the learning target and performing appropriate learning, the accuracy of diagnosis is improved.

(Examples 1 to 3)
This embodiment aims to improve the diagnostic accuracy by re-diagnosis. In particular, in the aspect of embodiment 1-1, this embodiment improves the diagnostic accuracy by performing measurement and diagnosis again when the diagnostic result is judged to be abnormal, and judging that it is truly abnormal when it is judged to be abnormal continuously. Basically, when an abnormality occurs in equipment, the abnormality continues to occur continuously, so it is judged to be abnormal when it is judged to be abnormal continuously.

The diagnostic process of Example 1-3 will be described with reference to the flowchart in FIG. 14. The process up to the diagnostic process is the same as in Example 1-1, so the description will be omitted. Below, the process after the diagnostic process of S118 (S131 to S134) will be described.

When an abnormality occurs in the rotating machine 3, the abnormality will generally continue to occur. Therefore, as shown in FIG. 15, once an abnormality is detected, the acceleration data is remeasured and rediagnosed, and if an abnormality is detected continuously, it is determined to be an abnormality. This makes it possible to reduce the number of false detections.

S131, S132: If an abnormality is determined in the diagnosis process of S118, the data measurement unit 6 performs a process of re-measuring the acceleration data. The re-measured acceleration data is stored in the memory unit 10.

S133: The diagnostic processing unit 14 performs re-diagnosis processing based on S118 using the re-measured acceleration data retrieved from the memory unit 10.

S134: If the re-diagnosis process in S133 again determines that there is an abnormality, the diagnosis result that there is an abnormality is transmitted by the diagnosis result transmission unit 15 to the monitoring and control system 7 in FIG. 2 or a cloud, etc. On the other hand, if there is a determination that there is a normality, the diagnosis result that there is a normality is transmitted to the monitoring and control system 7 or a cloud, etc. (not shown).

As described above, by re-diagnosing equipment abnormalities, the accuracy of diagnosis is improved by automatically determining whether the abnormality continues to occur and judging whether it is a genuine abnormality.

(Examples 1 to 4)
A trigger function is used to detect the operation of the rotating machine 3, and additional learning is performed on data after the start of operation. In particular, the sequential learning is started upon receiving a detection signal of the operating state of another object related to the rotating machine 3 or a manual trigger signal that commands the sequential learning. For example, if the load-side rotating equipment 5 is a blower as the other object, the vibration pattern changes when the equipment stops operating, and data after operation is provided for the sequential learning.

Figures 2 and 4 show the device configuration and an example of the configuration of a method for improving accuracy by additional learning of data after the start of operation using a trigger function. The abnormality diagnosis device 1 of this embodiment is equipped with a trigger measurement unit 8 that monitors the monitoring and control system data from the monitoring and control system 7 in Figure 2, determines the operation stop state of the rotating machine 3, and measures the acceleration data from the acceleration sensor 4. Note that it is also possible to determine the operation stop state of the rotating machine 3 by making a threshold judgment from the effective value of the acceleration data, etc., without using the monitoring and control system data.

The diagnostic process for Examples 1-4 will be explained with reference to the flowchart in Figure 16.

The diagnostic process of Example 1-4 is basically the same as the diagnostic process of Example 1-1, except that the trigger measurement process of S141 is performed in parallel with the learning period selection process of S111 to S115.

In the case of rotating equipment 5 such as a blower, the vibration pattern may change at the timing of operation shutdown as shown in FIG. 17. There is no problem if all data is learned sequentially, but depending on the performance of the CPU of the abnormality diagnosis device 1, it may be necessary to learn one piece of data per day or one piece of data per week sequentially. When learning one piece of data per day or one piece of data per week sequentially, the degree of deviation increases until new sequential learning is performed, as shown in FIG. 18.

Therefore, based on the "monitoring and control system data" from the monitoring and control system 7 in Figure 2 or the "acceleration data" from the acceleration sensor 4, a "trigger measurement" is performed to determine whether the rotating machine 3 has stopped operating and to measure the acceleration data when the rotating machine changes from a stopped state to an operating state. By sequentially learning the data obtained by this "trigger measurement," the number of false detections as shown in Figure 18 can be reduced. The "trigger measurement" is performed by a manual trigger as appropriate.

S141, S111 to S115: The rotating machine 3 is judged to have stopped operating based on the "monitoring and control system data" or "acceleration data", and the "acceleration data" is measured when the rotating machine 3 goes from a stopped state to an operating state. When using the "monitoring and control system data", the ON/OFF information of the equipment is used (S141). When using the "acceleration data", the effective acceleration value is calculated, and a threshold value is set to judge whether the rotating machine 3 has stopped operating, and the stopping of the equipment is judged (S111 to S115).

S116, S117: The sequential learning processing unit 13 performs sequential learning on the acceleration data obtained by the trigger measurement. In particular, the transition of the normal effective acceleration value when the operation of the diagnosis target is stopped is learned.

S118: The diagnostic processing unit 14 performs diagnosis using the learning result model after sequential learning.

S119: The diagnosis result transmission unit 15 executes a transmission process of the diagnosis result based on S118 to transmit the diagnosis result to the monitoring and control system 7, the cloud, etc.

According to the above-mentioned Examples 1 to 4, the diagnostic accuracy can be improved by additionally learning data after the start of operation using the trigger function. In particular, in the equipment abnormality diagnosis method, even for equipment whose vibration pattern changes when the operation is stopped, the diagnostic accuracy can be improved by learning data at the time of the operation stoppage.

[Embodiment 2]
19, an abnormality diagnosis device 1 measures and diagnoses vibration data of a rotating machine 3 and a rotating equipment 5. In particular, this embodiment narrows down the target data for diagnosis and learning from the measured acceleration data to only those suitable for diagnosis and learning, thereby improving the accuracy of diagnosis.

One or more acceleration sensors 4 are installed near the bearings of the rotating equipment 5, such as the rotating machine 3 or load to be measured. The number of acceleration sensors 4 is not limited to the four shown in the figure. A data measurement unit 6 that measures acceleration data is connected to the acceleration sensor 4. In addition, the data measurement unit 6 is connected to the abnormality diagnosis device 1. The abnormality diagnosis device 1 may be connected to a SCADA or cloud system to send the diagnosis results.

The abnormality diagnosis device 1 implements the following functional units shown in FIG. 20 through cooperation between the computer's hardware and software resources.

The data collection unit 16 collects the acceleration data output from the acceleration sensor 4 from the data measurement unit 6 and stores it in the memory unit 10.

The data filtering unit 17 analyzes the acceleration data retrieved from the memory unit 10 to determine whether the data is suitable for diagnosis and learning, and uses only suitable data for diagnosis and learning. Then, as described below, if the acceleration data is determined to be in a transitional period, it is determined to be unsuitable for diagnosis (and learning). Furthermore, the determination of suitability for diagnosis and learning is narrowed down to only a certain range of rotational frequencies.

The learning period selection unit 12 analyzes the acceleration data retrieved from the memory unit 10 and automatically selects the period during which the output of the acceleration sensor 4 is stable as the learning period.

The sequential learning processing unit 13 sequentially learns data from the learning period selected by the learning period selection unit 12. Note that since the timing of learning depends on the performance of the CPU of the abnormality diagnosis device 1, sequential learning may be performed by adding one piece of data per day, or one piece of data per week. In order to suppress annual fluctuations, sequential learning must be performed for at least one year, and after one year has passed, sequential learning may be continued or terminated.

The diagnostic processing unit 14 uses the learning result model learned by the sequential learning processing unit 13 to diagnose the state of the equipment from the current acceleration data.

The diagnosis result transmission unit 15 transmits the results of the diagnosis of the condition of the rotating machine 3 and the rotating equipment 5 by the diagnosis processing unit 14 to a monitoring control system, a cloud, etc.

(Example 2-1)
The diagnosis process of the embodiment 2-1 will be described with reference to the flowchart of FIG.

S211: An acceleration sensor 4 is installed on the bearings of the rotating machine 3 or rotating equipment 5 to be measured. The output signal of the acceleration sensor 4 is stored in the memory unit 10 as acceleration data of the rotating machine 3 via the data measurement unit 6 and the data collection unit 16. This measured acceleration data is stored in the memory unit 10 via the data measurement unit 6 and the data collection unit 16.

S212: The data filtering unit 17 narrows down the acceleration data retrieved from the memory unit 10 to data that can be used for diagnosis and learning. First, it is determined whether or not the vehicle is driving. This can be done by using the magnitude of the vibration effective value. If the vehicle is not driving, it is excluded from the diagnosis and learning targets. Next, it is determined whether or not the vehicle is in a transitional period using the method of S112 in the first embodiment. In summary, the measured acceleration data is divided into one-second intervals, each of which is frequency analyzed to estimate the rotational frequency for each divided element, and if there is variation in the rotational frequency for each element, it is determined to be in a transitional period. Furthermore, the estimated rotational frequency is used to narrow down the data. Only cases where the vehicle is not in a transitional period and the rotational frequency is within a certain range are considered to be usable for diagnosis and learning.

S213: If the transition period has been removed in S112 (No), the process returns to S211. On the other hand, if it is determined that the data can be used for diagnosis and learning (Yes), the process proceeds to the selection of a learning period in S214.

S214: If it is determined that the data can be used for diagnosis, the learning period selection unit 12 performs a learning period selection process. The learning period selection process is performed in the same manner as S114 in embodiment 1.

S215: If a learning period was not selected in S214 (No), the process returns to S211. On the other hand, if the learning period was selected (Yes), one-year sequential learning is performed in S216.

S216, S217: If a learning period is selected in S213, the sequential learning processing unit 13 performs sequential learning every day for one year. For learning, the method of Patent Documents 1-3 is used as in the first embodiment. Depending on the performance of the CPU of the abnormality diagnosis device 1, one piece of data may be learned per day, or one piece of data may be learned per week. After one year of sequential learning, the influence of annual variation disappears, so sequential learning is no longer necessary, but may be continued.

S218: The diagnostic processing unit 14 performs diagnostic processing to determine whether the current acceleration data is normal or abnormal using the model learned in S217. For the diagnostic processing, the method of Patent Documents 1-3 is used, as in the first embodiment.

S219: The diagnosis result transmission unit 15 executes a transmission process of the diagnosis result to transmit the diagnosis result obtained in S218 to a monitoring control system, a cloud, etc.

As described above, diagnosis and learning are limited to cases where the rotational frequency is within a certain range, so fluctuations in the diagnostic index value are suppressed.

(Example 2-2)
The rotation speed estimation process in S212 of the embodiment 2-1 is modified.

In S212 of Example 2-1, the estimated rotation frequency for each second of the measurement data (which is the same value unless it is a transitional period) is used, but in Example 2-2, the rotation frequency is estimated from the entire measurement data for one time (for example, 5 seconds) separately from removing the transitional period. This allows the rotation frequency to be estimated with finer accuracy than in Example 2-1, and the diagnosis target can be narrowed down to more specific conditions.

(Example 2-3)
In the embodiment 2-1 and the embodiment 2-2, how to determine the range of rotational frequencies to be narrowed down as a diagnosis target depends on human knowledge. In the present embodiment, a method of determining the range of rotational frequencies to be narrowed down is provided.

The rotational frequency range narrowed down in the process of S212 in Example 2-1 is as follows:

First, the rotational frequency range to be narrowed down is the entire range. In other words, the range from the minimum rotational speed to the maximum rotational speed is the subject of diagnosis (and learning). At this time, the calculated estimated rotational frequency is recorded in the memory unit 10.

If the diagnostic results show a large variation after a certain period of time (e.g., one month), the estimated rotational frequencies recorded in the memory unit 10 are statistically processed to narrow down the most frequent rotational frequencies to be diagnosed and learned. At this time, if the frequency of the subjects for diagnosis and learning is not high enough (e.g., the number is less than 1/4 of the total number of operating data), the range of rotational frequencies is expanded both upward and downward to increase the number of subjects for diagnosis. Note that if the diagnostic results show a large variation, there is no need to narrow down the rotational frequencies, as the variation in rotational frequency does not have a large effect on the diagnostic results.

(Example 2-4)
In the embodiment 2-3, the decision as to whether or not to start narrowing down the rotation frequency and the execution of the statistical processing are performed manually, whereas in the embodiment 2-4, the decision and the statistical processing are incorporated into the system.

As shown in FIG. 22, the abnormality diagnosis device 1 of Example 2-4 implements the following functional units through cooperation between the hardware and software resources of a computer.

The transitional period removal unit 11, the learning period selection unit 12, the sequential learning processing unit 13, the diagnostic processing unit 14, and the diagnostic result transmission unit 15 are similar to the transitional period removal unit 11, the learning period selection unit 12, the sequential learning processing unit 13, the diagnostic processing unit 14, and the diagnostic result transmission unit 15 of the first embodiment.

The data filtering unit 17 filters out the acceleration data retrieved from the memory unit 10 and records the rotational frequency when it is not in a transitional period in the memory unit 10. In particular, the data filtering unit 17 excludes acceleration data that is determined to be in a transitional period of the operating state of the diagnosis target based on the peak fluctuation in the rotational frequency band obtained from the acceleration data of the diagnosis target from the sequential learning. In addition, the data filtering unit 17 obtains filtering parameters (such as the rotational frequency range) from the memory unit 10.

When a certain amount of acceleration data (e.g., one month's worth) has been accumulated, the parameter update unit 18 performs statistical processing on the rotational frequencies to narrow down the rotational frequency range to the most frequent rotational frequency, and updates the narrowing down parameters in the memory unit 10.

The aperture parameter update process is explained using the flowchart in Figure 23.

The parameter update unit 18 is executed separately from the normal diagnostic and learning processes after a certain period of time (e.g., one month) has passed since the start of operation.

S241: Calculate the rotation frequency histogram.

S242: Set the most frequent rotation frequency to the rotation frequency range.

S243: If the frequency of the rotation frequency range is 1/4 or more of the total, end; if not, proceed to S244.

S244: The higher frequency above or below the current range is added to the rotation frequency range, and S243 is executed again.

Figure 24 shows a distribution diagram of the rotational frequency ratio for the entire period and the first month, estimated from data measured over a period of just over two years for a certain piece of equipment. The figure shows the distribution of rotational frequencies for a period of just over two years for a certain piece of equipment.

As can be seen, the rotational frequency distribution is roughly accurate within the first month, and the frequency of the most frequent rotational frequency is about 30%. Even if the rotational frequencies to be diagnosed are narrowed down, about one-third of the operating time is subject to diagnosis. In this example, the operating time ratio is about 50%, and 3 to 4 diagnoses can be expected per day.

Figures 25 and 26 show graphs of the results of diagnosis with no rotational frequency limit and with a rotational frequency limit of 19 Hz, based on data measured over a period of just over two years for the same equipment as above. As can be seen, in Figure 25, where there is no rotational frequency limit, index values that frequently exceed the abnormal threshold occur, but in Figure 26, where the rotational frequency is limited to 19 Hz, the index values do not show any abnormal values except in the early stages when learning has not progressed. They are below the caution line. The frequent abnormal values shown in Figure 25 are because the index values fluctuate greatly with changes in the rotation speed. By narrowing down the rotational frequencies diagnosed in Figure 26, false positives due to diagnostic abnormalities such as those in Figure 25 can be avoided.

[Embodiment 3]
When the abnormality diagnosis methods of Patent Documents 1-3 and the like are applied to a variable-speed rotating machine, the rotation frequency changes for each measurement of sensor data, so that even normal data may vary and cause erroneous detection. FIG. 27 shows an example of a frequency analysis result when the rotation frequency of a variable-speed rotating machine from 16 Hz to 20 Hz is 16 Hz and 20 Hz. When the rotation frequency is 16 Hz and 20 Hz, there may be a shift in the position of the peak frequency or the frequency intensity may differ. Learning such data collectively may result in a failure to perform a normal diagnosis.

In contrast, in the second embodiment, a distribution of rotational frequencies is created and only data of rotational frequencies with high frequency is learned, or only data in the range of rotational frequencies with high frequency is learned, thereby improving the accuracy of diagnosis.

However, narrowing down the range to rotational frequencies with high frequency also limits the frequency at which diagnosis is performed, so it is necessary to widen the range of rotational frequencies. There is a problem in that it is difficult to automatically set the range of rotational frequencies required for accurate diagnosis by simply setting the range of rotational frequencies with high frequency.

The abnormality diagnosis device 1a of this embodiment shown in FIG. 28 automatically sets the range of rotational frequencies in the abnormality diagnosis method of embodiment 1, taking into account the accuracy of the diagnosis method.

The abnormality diagnosis device 1a illustrated in Figs. 28 and 29 is similar to the abnormality diagnosis device 1 of embodiment 1 except that a data filtering unit 17 is added, and the following functional units are implemented by cooperation between the hardware resources and software resources of a computer. Note that the data filtering unit 17 has different processing contents from the data filtering unit 17 of embodiment 2.

The data measurement unit 6 stores the acceleration data output from the acceleration sensor 4 in the memory unit 10.

The data filtering unit 17 automatically determines whether the acceleration data retrieved from the memory unit 10 can be used for learning or diagnosis, and narrows down the acceleration data.

The transient period removal unit 11 determines whether the operating state of the rotating machine 3 is in a transient period based on the acceleration data retrieved from the memory unit 10, and if so, excludes it from the targets of sequential learning and diagnosis.

The learning period selection unit 12 automatically selects a period during which the sensor output is stable as the learning period based on the acceleration data retrieved from the memory unit 10.

The sequential learning processing unit 13 sequentially learns from the selected learning period.

The diagnostic processing unit 14 diagnoses the state of the equipment from the current acceleration data based on the learning result model obtained by the learning.

The diagnosis result transmission unit 15 transmits the results of the diagnosis of the equipment condition by the diagnosis processing unit 14 to a monitoring control system, cloud, etc.

(Example 3-1)
The diagnosis process of Example 3-1 will be described with reference to the flowchart of Fig. 30. The process flow of this embodiment is the same as the process flow of the first embodiment except for S302.

S301: The acceleration sensor 4 is installed on the bearing of the rotating machine 3 to be measured, and the acceleration data of the rotating machine 3 is measured by the data measurement unit 6. This measured acceleration data is stored in the memory unit 10.

S302: To perform a stable diagnosis, the data selection unit 17 narrows down whether the acceleration data in the memory unit 10 is to be used for learning or diagnosis based on the rotation frequency. The basic idea is as follows.

As shown in Fig. 27, if learning and diagnosis are performed on data in which the results of frequency analysis for each rotation frequency are different, stable diagnosis cannot be performed. Therefore, frequency analysis is performed for each rotation frequency, the correlation of the frequency analysis results is calculated, and data in the range of rotation frequencies with a high correlation value is used. As an example of processing, consider a variable-speed rotating machine 3 whose rotation frequency varies from 16 Hz to 20 Hz.
(1) As shown in Fig. 31, a frequency analysis is performed on the measurement data, and the data is classified by rotation frequency. As for the rotation frequency, the frequency that peaks between 16 Hz and 20 Hz in the frequency analysis result is determined as the rotation frequency.
(2) Calculate the average value of the frequency analysis results for each rotation frequency.
(3) Calculate the correlation value of each of the average data of the frequency analysis results for each rotation frequency. An example of the calculation results of the correlation values is shown in Table 1.
(4) In Table 1, the minimum correlation value is calculated, and the processes from S313 onward are performed for data in the range of rotational frequencies in which the minimum correlation value is 0.95 or more. The threshold value for the minimum correlation value may be determined arbitrarily.

S303: The transient period removal unit 11 determines whether the measured acceleration data is in a transient period. The method for determining whether the measured acceleration data is in a transient period is the same as that in embodiment 1.

S304: If it is determined in S303 that the period is not a transitional period, the learning period selection unit 12 performs a learning period selection process. The selection method in embodiment 1 is applied to the learning period selection method.

S305: If a learning period was not selected in S304 (No), the process returns to S301. On the other hand, if the learning period was selected (Yes), one-year sequential learning is performed in S306.

S306: If a learning period is selected in S305, the sequential learning processing unit 13 performs sequential learning every day for one year. In the sequential learning, the method of, for example, Patent Documents 1-3 is applied as in the first embodiment. Once one year's worth of sequential learning has been performed, the process proceeds to the diagnosis process of S118.

S307: If one year's worth of learning has not been performed, sequential learning is performed. If one year's worth of learning has been performed, the diagnosis process of S118 is performed. Depending on the performance of the CPU of the abnormality diagnosis device 1 that performs sequential learning, one piece of data may be learned per day, or one piece of data per week. After one year of learning, the influence of annual variation disappears, so sequential learning is no longer necessary, but it may be continued.

S308: The diagnostic processing unit 14 performs diagnostic processing to determine whether the current acceleration data is normal or abnormal using the model learned in S307. For the diagnostic processing, the methods described in Patent Documents 1-3 may be used.

S309: The diagnosis result transmission unit 15 executes a process for transmitting the diagnosis result obtained in S308 to a monitoring control system, a cloud, etc.

As described above, when diagnosing a variable speed rotating machine, the accuracy of the diagnosis can be improved by automatically setting the range of rotational frequencies for stable diagnosis. In addition, the number of data to be excluded can be minimized by narrowing down the data to be diagnosed. In particular, it is possible to automatically narrow down the data of rotational frequencies that allow stable diagnosis from the sensor data of the acceleration of the variable speed rotating machine for which abnormality diagnosis is desired, and it is possible to diagnose the abnormal state of the equipment with higher accuracy, even compared to embodiment 2.

[Embodiment 4]
In the conventional anomaly diagnosis methods disclosed in Patent Documents 1 to 3, first, a diagnostic criterion (a transformation matrix and a normalization coefficient) is calculated from a plurality of normal time series data prepared in advance in a preparation phase.

In contrast, the abnormality diagnosis device 1 of the above-mentioned embodiment 2 adds a part of the time series data used for diagnosis to new normal time series data used for learning and re-executes the preparation phase (Figs. 20 and 21) in order to improve the accuracy of diagnosis by tracking normal states that can change over time. In the diagnosis process of Fig. 21, a sequential learning process is performed before diagnosis (S217, S218), and since it is difficult to separate diagnosis and learning, the learning stage (preparation phase) is also performed by the abnormality diagnosis device.

The heaviest process in constant-Q diagnosis (processing requiring a large amount of calculations and memory usage) is the principal component analysis in the preparation phase. The constant-Q transformation of Patent Documents 1-3, which is executed in both the preparation and diagnosis phases, is also a fairly heavy calculation process, and these documents propose a method to speed it up. The amount of calculation for the constant-Q transformation is determined by the length of the time series data, so when dealing with time series data of the same length each time, as in these documents, the amount of calculation in the diagnosis phase is roughly the same each time, and the amount of calculation in the preparation phase is roughly proportional to the number of time series data used.

Principal component analysis requires a quadratic amount of calculations or more in relation to the number of multivariate samples, and memory usage also increases. The number of multivariate samples is proportional to the number of time series data used in the preparation phase. For this reason, executing the preparation phase with a relatively large number of time series data requires enormous computer resources (amount of calculations and memory). For this reason, while it is possible to execute the preparation phase using a large amount of time series data without any problems in an environment with ample computer resources, such as a high-performance PC or cloud server, in an anomaly diagnosis device that has limited computer resources because it is placed near the field equipment that measures the time series data, it becomes difficult to execute the preparation phase as the number of target time series data increases.

For the diagnosis of Patent Documents 1-3, there is no problem if the preparation phase is carried out in an environment where sufficient computer resources can be secured in advance, and the anomaly diagnosis device is operated based on the diagnostic criteria.

On the other hand, in the abnormality diagnosis device 1 of embodiment 2, if a sequential learning process including a preparation phase is to be performed, excessive computer resources (CPU, memory) are required for the abnormality diagnosis device. Such excessive computer resources have issues such as size, heat dissipation performance, power consumption, and price, and it is difficult to equip all abnormality diagnosis devices with them.

Therefore, the equipment anomaly diagnosis system 2 of embodiment 4 shown in FIG. 33 changes the order of the sequential learning process and separates the re-learning process in the sequential learning and constant Q diagnosis of embodiment 2, and performs the re-learning process, including the preparation phase, which is a heavy process, in a learning device 21 separate from the anomaly diagnosis device 1.

The equipment abnormality diagnosis system 2 of this embodiment performs abnormality diagnosis of a rotating machine 3. The rotating machine 3 is connected to rotating equipment 5 such as a pump, turbo blower, water wheel, etc. An acceleration sensor 4 is attached to the rotating machine 3. Acceleration data of the rotating machine 3 detected by the acceleration sensor 4 is output to the abnormality diagnosis device 1 via a data measurement unit 6. The data measurement unit 6 receives the output signal of the acceleration sensor 4 as acceleration data of the rotating machine 3 and outputs it to the abnormality diagnosis device 1. The abnormality diagnosis device 1 selects the acceleration data and performs diagnosis and learning. At this time, the learning process delegates re-learning processing to the learning device 21. The abnormality diagnosis device 1 may also be connected to an external system 22 such as a display device, a monitoring and control system, or a cloud system as a destination for sending the diagnosis results.

The abnormality diagnosis device 1 implements the following functions shown in FIG. 34 through the cooperation of computer hardware and software resources.

The data collection unit 16 collects the acceleration data output from the acceleration sensor 4 from the data measurement unit 6 and stores it in the memory unit 10.

The data filtering unit 17 analyzes the acceleration data retrieved from the memory unit 10 to determine whether the data is suitable for diagnosis and learning, and uses only suitable data for diagnosis and learning. Then, as described below, if the acceleration data is determined to be not in operation or is in a transitional period, it is determined to be unsuitable for diagnosis and learning. Furthermore, the determination of suitability for diagnosis and learning may be narrowed down to only a certain range of rotational frequencies.

The diagnostic processing unit 14 diagnoses the acceleration data that the data filtering unit 17 has determined to be suitable for diagnosis and learning. At this time, if there is not enough data available for learning and the sequential learning processing unit is unable to calculate diagnostic reference data, the data is deemed to be outside the scope of diagnosis and no diagnosis result is given. In particular, the data is not initially subject to diagnosis because it has not yet been learned.

The diagnosis result transmission unit 15 transmits the results of the diagnosis of the condition of the rotating machine 3 and the rotating equipment 5 by the diagnosis processing unit 14 to the external system 9.

The learning period selection unit 12 analyzes the acceleration data retrieved from the memory unit 10 and automatically selects the period during which the output of the acceleration sensor 4 is stable as the learning period.

The sequential learning processing unit 13 sequentially extracts learning data from the learning period selected by the learning period selection unit 12. At this time, only acceleration data that is determined by the data narrowing unit 17 to be suitable for diagnosis and learning is the target. Acceleration data with abnormal diagnosis results may be excluded. Since the number of learning targets and the timing and frequency of learning depend on the calculation performance of the learning device 21, sequential learning may be performed by adding one data item per day, or one data item per week. In order to suppress yearly fluctuations, sequential learning must be performed for at least one year, and after one year has passed, sequential learning may be continued or terminated. When learning data is added, the actual learning process is performed by transmitting the acceleration data to the learning device 21 and the learning device 21 performs the process.

The diagnosis process of the first embodiment will be described with reference to the flowcharts of FIGS.
This diagnostic process is executed periodically or in response to a trigger.

S211: An acceleration sensor 4 is installed on the bearings of the rotating machine 3 or rotating equipment 5 to be measured. The output signal of the acceleration sensor 4 is stored in the memory unit 10 as acceleration data of the rotating machine 3 via the data measurement unit 6 and the data collection unit 16.

S212: The data filtering unit 17 narrows down the acceleration data retrieved from the memory unit 10 to data that can be used for diagnosis and learning. First, it is determined whether or not the vehicle is being driven. This can be done by using the magnitude of the vibration effective value. If the vehicle is not being driven, it is excluded from the diagnosis and learning targets. Next, it is determined whether or not the vehicle is in a transitional period using the method of embodiment 1. In summary, the measured acceleration data is divided into one-second intervals, each of which is frequency analyzed to estimate the rotational frequency for each divided element, and if there is variation in the rotational frequency for each element, it is determined to be in a transitional period. Furthermore, the estimated rotational frequency is used to narrow down the data. Only cases where the vehicle is not in a transitional period and the rotational frequency is within a certain range are considered to be usable for diagnosis and learning.

S213: If the data was removed in S212 (no), the current diagnostic process ends. On the other hand, if the data is determined to be usable for diagnosis and learning (yes), the process moves to S451 to determine the diagnostic criteria.

S251: If there is a diagnostic criterion (yes), the diagnostic processing unit 14 executes the diagnostic process S418. If there is no diagnostic criterion (no), a diagnostic result of "No diagnosis" S452 is created, indicating that the diagnostic target is not included. The diagnostic criterion is created in the preparation phase of Patent Documents 1-3. In this embodiment, the diagnostic criterion is created in the sequential learning process S217. Since there is a possibility that the diagnostic criterion does not exist before executing the sequential learning process S217, it is determined here.

S252: If there are no diagnostic criteria, the diagnostic processing unit 14 creates a diagnostic result indicating that the item is not subject to diagnosis.

S218: If there is a diagnostic criterion, the diagnostic processing unit 14 performs diagnostic processing to determine whether the current acceleration data is normal or abnormal. This diagnostic processing is the processing of the diagnostic phase described in Patent Documents 1 and 2.

S219: The diagnosis result transmission unit 15 executes a diagnosis result transmission process to transmit the diagnosis result obtained in S218 or S252 to the external system 22.

S212: The learning period selection unit 12 performs a learning period selection process. The learning period selection process is performed in the same manner as in embodiment 2.

S215: If no learning period was selected in S214 (no), this diagnostic process ends. On the other hand, if the learning period was selected (yes), proceed to S216 and learning will be performed sequentially up to one year.

S216: If a learning period is selected in S213, the sequential learning processing unit 13 performs sequential learning every day for one year. After one year of sequential learning, the influence of annual variation disappears, so sequential learning does not need to be performed, but it may be continued.

S217: The sequential learning processing unit 13 performs sequential learning every day for one year. The sequential learning process in the sequential learning processing unit 13 is described in the flowchart of FIG. 36.

S253: In the sequential learning process, first, learning data is selected. Basically, all data determined in S212 to be usable for diagnosis and learning may be used as learning data, but depending on the CPU performance of the learning device 21, the learning data may be thinned out, such as learning one piece of data per day, or one piece of data per week. Also, when there is a certain amount of learning data, data with abnormal diagnosis results may also be excluded from the learning data.

S254: If no new learning data is selected by S253 (no), the current sequential learning process ends. On the other hand, if new learning data is selected (yes), a remote learning request is made (S255).

S255: If new learning data is selected, the sequential learning processing unit 13 sends the new learning data to the learning device 21 and requests remote learning.

S256: In the remote learning process, when the learning device 21 receives new learning data from the sequential learning processing unit 13, it recalculates the diagnostic criteria together with the learning data that has already been stored. This process uses, for example, the preparation phase methods described in Patent Documents 1-3.

S257: The sequential learning processing unit 13 waits for the completion of the remote learning process in the learning device 21 and receives the returned diagnostic criteria.

S258: The sequential learning processing unit 13 stores the diagnostic criteria returned from the remote learning process in the memory unit 10 so that they can be used for future diagnoses.

Depending on the CPU performance of the learning device 21, the remote learning process may take a longer calculation time than the period of the diagnostic process. In this case, the learning data selected in the learning data selection process S253 is appropriately thinned out so that the next remote learning process is not called before the remote learning process is completed. In this case, the diagnostic criteria cannot be updated in time for the next diagnostic process, so the previous diagnostic criteria are used until the diagnostic criteria are updated. Furthermore, the update of the diagnostic criteria is appropriately controlled by mutual exclusion so that inconsistent diagnostic criteria are not used in the diagnostic process S218.

According to the present embodiment described above, the order of the diagnostic process and the learning process is rearranged in sequential learning and constant Q diagnosis, and the heavy learning process is provided to a high-speed learning device 21 separate from the abnormality diagnosis device 1, thereby further speeding up the learning process.

[Embodiment 5]
37 reuses multivariate samples (constant-Q transformation results) created during diagnosis in the manner of embodiment 4. However, it is assumed that the learning device 21 is placed in a remote location, such as on a cloud system, far away from the abnormality diagnosis device 1, and the amount of data to be transmitted needs to be reduced due to the network bandwidth connecting the abnormality diagnosis device 1 and the learning device 21.

The diagnostic processing unit 14 performs diagnosis in the same manner as in embodiment 4, but also stores the multivariate samples of the constant Q transformation results calculated during the diagnostic processing in the memory unit 10 together with the diagnostic result data.

The sequential learning processing unit 13 reads out the acceleration data and the corresponding multivariate samples at the same time. When learning data is selected, the data sent to the learning device 21 is not the acceleration data but the corresponding multivariate samples.

When performing remote learning processing, the learning device 21 omits the constant Q transformation and performs the preparatory phase processing of Patent Documents 1-3 from the multivariate samples.

The diagnostic process of this embodiment is also based on the flowcharts of Figures 35 and 36. This diagnostic process is the same as that of embodiment 4 except for the diagnostic processes (S252, S218, S255, and S256).

S252: If there are no diagnostic criteria, the diagnostic processing unit 14 creates a diagnostic result that the diagnosis is not applicable. Also, even if there are no diagnostic criteria, the constant Q transformation is performed and the multivariate sample is stored in the storage unit 10.

S218: If there is a diagnostic criterion, the diagnostic processing unit 14 performs a diagnosis in the same manner as in embodiment 4, and stores in the memory unit 10 a multivariate sample of the constant Q transformation result calculated during the diagnosis.

S255: When new learning data is selected, the sequential learning processing unit 13 sends a multivariate sample corresponding to the new learning data to the learning device 21 and requests remote learning.

S256: In the remote learning process, the learning device 21 sequentially accepts new multivariate samples of learning data from the learning processing unit 13, stores them in the learning device 21, and recalculates the diagnostic criteria together with the learning data (multivariate samples) that have already been stored. This process uses, for example, the preparation phase methods described in Patent Documents 1-3. However, since there are multivariate samples that have already been constant-Q transformed, the constant-Q transformation is omitted and processing begins with principal component analysis.

Therefore, according to the diagnostic process of this embodiment, the data used for learning is multivariate samples created at the time of diagnosis from acceleration data, and the constant Q transformation during learning is omitted, speeding up the learning process, and the transmission load of the learning data is reduced by making the size of the multivariate samples smaller than the original acceleration data.

The effect of reducing the amount of data transmitted in embodiment 5 is calculated as follows:

Acceleration data recorded at 51.2kHz for 5 seconds is approximately 2MB (51200 x 5 x 8 = 2048000) with each value being a double-precision floating-point number (8 bytes).

On the other hand, there are 63 frequency bands from 6.25 Hz to 8064 Hz in the 1/6 octave band, and if 100 samples are generated from 5 seconds of samples, the multivariate samples resulting from the constant Q conversion will be about 50 kB (63 x 100 x 8 = 50,400) (values are double-precision floating-point numbers).

For this reason, by converting the acceleration data into multivariate samples, the data size is reduced to 1/40. For example, even if the measurement value of the acceleration sensor can be obtained as a 2-byte integer and the acceleration data is reduced to 1/4 of the size mentioned above, the multivariate sample resulting from the constant Q conversion will be 1/10 of that size. By converting the learning data into multivariate samples in this way and then transmitting them to the learning device 21, the effect of reducing transmission costs is enhanced.

[Embodiment 6]
In the diagnostic process of this embodiment, in the aspects of the fourth and fifth embodiments, the learning device 21 limits only the level at which the target data affects re-learning by multi-level constant Q diagnosis.

As the learning device 21, an abnormality diagnosis device that executes the preparation phase and diagnosis phase disclosed in Patent Document 3 is applied. That is, as shown in FIG. 38, the learning device 21 is equipped with a parameter creation unit 100 that executes the preparation phase and a diagnosis unit 200 that executes the diagnosis phase, and executes multi-level constant Q diagnosis.

The parameter creation unit 100 includes a constant-Q conversion unit 101, a principal component analysis unit 102, a statistical value calculation unit 103, and a normalization unit 104. The constant-Q conversion unit 101 performs constant-Q conversion on time-series data D1 collected in advance by measuring (measuring) a diagnosis target for abnormality diagnosis, to create multivariate samples (learning samples) to be used for learning. The principal component analysis unit 102 performs principal component analysis on the multivariate samples to obtain a transformation matrix and calculates a principal component score for each multivariate sample. The statistical value calculation unit 103 obtains a statistical quantity that serves as an index of the multivariate sample based on the principal component scores. The normalization unit 104 normalizes the statistical quantity to obtain the degree of anomaly.

The diagnosis unit 200 includes a constant-Q transformation unit 201, a principal component calculation unit 202, a second statistical value calculation unit 203, and a normalization unit 204. The constant-Q transformation unit 201 performs constant-Q transformation on newly measured time series data D2 to create a multivariate sample. The principal component calculation unit 202 obtains a principal component score based on the transformation matrix obtained from the multivariate sample by the principal component analysis unit 102. The second statistical value calculation unit 203 obtains a statistic based on the principal component score. The normalization unit 204 normalizes the statistic by the normalization coefficient used in the normalization unit 104 to calculate the degree of anomaly. The diagnosis unit 200 then diagnoses an anomaly in the target for anomaly diagnosis based on the degree of anomaly.

The diagnostic process of this embodiment is similar to that of the fourth and fifth embodiments based on the flowcharts of FIGS. 35 and 36, except for the remote learning process (S256). The remote learning process (S256) of this embodiment will be described below.

S256: Multi-level constant-Q diagnosis creates diagnostic criteria by performing principal component analysis on multivariate samples of the same or adjacent levels for each level of time-series data. In this case, when a new learning data is added, the diagnostic criteria that are updated are only those for the same or several adjacent levels.

Therefore, the re-learning process can be reduced by limiting the diagnostic criteria to be re-learned when new training data is added to diagnostic criteria of the same or a few adjacent levels. In particular, when performing multi-level constant-Q diagnosis, the re-learning process when new training data is selected can be limited to the range of adjacent levels where the new training data affects the diagnostic criteria, making re-learning more efficient.

Reference Signs List 1, 1a... Abnormality diagnosis device 2... Equipment abnormality diagnosis system 3... Rotating machine 4... Acceleration sensor 5... Rotating equipment 6... Data measurement unit 7... Monitoring and control system 8... Trigger measurement unit 10... Memory unit 11... Transient period removal unit 12... Learning period selection unit 13... Sequential learning processing unit 14... Diagnosis processing unit 15... Diagnosis result transmission unit 16... Data collection unit 17... Data narrowing down unit 18... Parameter update unit 21... Learning device, 100... Parameter creation unit, 101... Constant Q conversion unit, 102... Principal component analysis unit, 103... Statistical value calculation unit, 104... Normalization unit, 200... Diagnosis unit, 201... Constant Q conversion unit, 202... Principal component calculation unit, 203... Second statistical value calculation unit, 204... Normalization unit 22... External system

Claims (11)

An abnormality diagnosis device that sequentially learns an operating state of a diagnosis target and diagnoses abnormalities,
a transient period elimination unit that excludes acceleration data determined to be in a transient period of an operating state of the diagnosis object based on a fluctuation in acceleration amplitude obtained from the acceleration data of the diagnosis object from the sequential learning;
The diagnosis target is a variable speed rotating machine, and a specific rotation frequency set based on acceleration data of the rotating machine is used for the sequential learning and the diagnosis.
An abnormality diagnosis device comprising: An abnormality diagnosis device that sequentially learns an operating state of a diagnosis target and diagnoses abnormalities,
an abnormality diagnosis device comprising: a data filtering unit that excludes acceleration data determined to be in a transitional period of the operating state of the object to be diagnosed based on peak fluctuations in a rotational frequency band obtained from the acceleration data of the object to be diagnosed from the sequential learning. 3. The abnormality diagnosis device according to claim 1, further comprising a learning period selection unit that selects a time period during which a standard deviation of a moving average of the acceleration effective value of the acceleration data converges within a predetermined range as the sequential learning period. The abnormality diagnosis device according to any one of claims 1 to 3, characterized in that it is provided with a sequential learning processing unit that starts the sequential learning upon receiving a detection signal of the operating state of another object related to the diagnosis object or a manual trigger signal that commands the sequential learning. The abnormality diagnosis device according to claim 3, further comprising a sequential learning processing unit that learns the normal transition of the effective acceleration value when the operation of the diagnosis target is stopped. 3. The abnormality diagnosis device according to claim 2 , wherein the diagnosis target is a variable speed rotating machine, and a specific rotation frequency set based on acceleration data of the rotating machine is provided for the sequential learning and the diagnosis. An abnormality diagnosis system that sequentially learns an operating state of a diagnosis target and diagnoses abnormalities,
The abnormality diagnosis device according to claim 1 , which performs the sequential learning and the diagnosis;
an abnormality diagnosis system comprising: a learning device that re-learns the diagnosis criteria based on the multivariate samples of the acceleration data obtained by the sequential learning and existing multivariate samples of the acceleration data. The abnormality diagnosis system according to claim 7, characterized in that the learning device performs the re-learning without performing constant Q conversion of the time series data of the acceleration data. The abnormality diagnosis system according to claim 7, characterized in that the learning device performs the re-learning based on multivariate samples of the same or close levels of the existing acceleration data obtained by constant Q transformation of the time series data of the acceleration data, and multivariate samples of the same or close levels of the newly measured acceleration data obtained by constant Q transformation of the time series data of the acceleration data. An abnormality diagnosis method for sequentially learning an operating state of a diagnosis target and diagnosing an abnormality, comprising:
an abnormality diagnosis method comprising a step of excluding acceleration data determined to be in a transitional period of the operating state of the object to be diagnosed based on peak fluctuations in a rotational frequency band obtained from the acceleration data of the object to be diagnosed from the sequential learning. The abnormality diagnosis method according to claim 10, further comprising a step of re-learning the diagnostic criteria based on the multivariate samples of the acceleration data obtained by the sequential learning and existing multivariate samples of the acceleration data.

Images