JPH04106429A - Abnormality diagnosing apparatus for rotary machine - Google Patents

Abstract

PURPOSE:To obtain an apparatus which allows highly accurate diagnosis of a cause of abnormality in a rotary machine by inputting a vibration analysis data with limited variations per lot of a rotary machine into a neural network (NNW). CONSTITUTION:An NNW2 learns a corresponding relationship between an input data from a screw compressor 10 and causes of abnormality in the compressor 10 corresponding to the input data using study algorithm of a back propagation type to determine a link weight of a linking section 6 of the NNW2 and outputs an output data (y) involving a cause of abnormality based on the link weight at a diagnosis. In conception, the NNW2 is of a three-layer structure of such a type. The link weight determined in the learning of the NNW2 is stored into a memory 3. Then, an arithmetic control section 4 computes a data pertaining to vibration from the compressor 10 as a data concerning a ratio of change with the data in the normal operation of the compressor 10 to be inputted into the NNW2. At the same time, the cause of the abnormality of the compressor 10 is diagnosed based on the output data from the NNW2 to be outputted to the compressor 10.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention provides (1) an abnormality diagnosis for diagnosing the cause of an abnormality in a rotating machine, such as a screw compressor that compresses fluid, based on vibration analysis data from the rotating machine; The present invention relates to an abnormality diagnosing apparatus for a rotating machine, and in particular to an abnormality diagnosing apparatus for a rotating machine that applies a backpropagation type neural network to diagnose the cause of the abnormality.

[Prior art]

A screw compressor 10, which is an example of the above-mentioned rotating machine, is shown in FIG. 7. This screw compressor 1110 is
An oil-free screw compressor for compressing air, which includes a suction casing 21. Rotor casing 22. A pair of male and female rotors 24 and 25 that mesh with each other in a non-contact manner are housed in a casing 20 made of an end cover 23, and each rotor shaft 26, 27 is connected to the casing 2.
Shaft seal 28 and bearing 3 which serve as a bearing part provided in 0
5.38 or shaft seal 29 and bearing 36.39, or shaft seal 30 and bearings 34, 36, or shaft seal 31 and bearings 35, 364, respectively, are rotatably supported. The rotor shaft 26 and the rotor shaft 27 are synchronously interlocked by a timing gear 3233 that is fixed to the shaft end on the end cover 23 side and meshes with each other. Further, a drive gear 37 for transmitting power from a drive source (not shown) is fixed to the other end side of the timing gear 32 of the rotor shaft 26.

Since such a screw compressor 10 is an oil-free type, the pair of male and female rotors 24 and 25 do not come into contact with each other and rotate synchronously while maintaining a small gap between them. ing. Therefore, the gap between the timing gears 32 and 33 is set to be smaller than the minute gap between the respective rotors, taking into account the thermal expansion of the female rotors 24 and 25 during operation.

Therefore, when the rotor shaft 26 of the male rotor 24 is rotationally driven by the drive source, the pair of male and female rotors 24.
25 rotate synchronously. As a result, the suction air sucked from the suction casing 2I side is guided into the suction space formed by the teeth of the rotors 24 and 25 in the rotor casing 22. At this time, the suction space formed by each rotor 2425 gradually decreases as each rotor 24.25 rotates, so the air sealed in the suction space is gradually compressed and The air is discharged as high-pressure air from a discharge port (not shown) of the casing 220.

In such a screw compressor 10, each of the rotors 24 and 25 needs to be rotated at a high speed in order to obtain preferable compression performance, and therefore each of the rotors 24 and 25 is likely to generate shaft vibration. Ta. This may lead to abnormalities in the bearings and timing gears 32 and 33. Furthermore, although each of the rotors 24 and 25 is constructed taking thermal expansion into consideration, for example, if the machining accuracy of each of the rotors 24 and 25 is poor, or if the setting in the rotor casing 22 is incorrect, the operation may be interrupted. The heat of compression caused these to reach high temperatures, and their thermal expansion caused them to come into contact with each other.

Therefore, the screw compressor 10 is equipped with a bearing monitor that detects abnormalities in the bearings, or a skilled operator places a listening rod on the screw compressor 10 to detect abnormal sounds and determine the location of the abnormality. Was. however.

If the above-mentioned method uses a listening rod to detect abnormal sounds, it is necessary to train skilled operators with deep skill and experience, and it is also necessary to have these skilled operators on hand at all times.

By the way, as an example of the abnormality diagnosis method that requires skill and experience as described above, it is possible to apply a neural network in which a learning step is executed based on a backpropagation type learning algorithm. This neural network is conceptually a multi-layered network, and during learning, vibration analysis data for each cause of abnormality in the screw compressor 10 is input to the input layer of the network, and output data related to the cause of the abnormality at this time is input to the input layer of the network. The connection weights of this neural network are changed and determined in order to reduce the difference between the output data 3 output from the output layer and the desired output data (teacher data) related to the cause of the abnormality. This determined connection weight is.

For example, stored in memory. The above-mentioned vibration analysis data refers to waveform data related to vibrations from a vibration sensor (not shown) attached to the screw compressor 10, which is converted from a time domain to a frequency domain using a waveform analyzer, etc. Value 2 means the value calculated as the probability density in the amplitude direction in the time domain or the value obtained by processing the predetermined waveform change while keeping the time rn' FA. When diagnosing an abnormality, new vibration analysis data regarding the vibration of the screw compressor IO is input to the neural network 7. Then, this neural network calculates the vibration analysis data at this time based on the network connection weights stored in the memory, and outputs output data indicating the cause of the abnormality of the screw compressor 10 from the output layer. As a result, the screw compressor 10
It is now possible to accurately diagnose the failure location that is the cause of the abnormality.

[Problems to be Solved by the Invention] In diagnosing the causes of abnormalities in the screw compressor 10 as described above, trend management of the vibration analysis data is often performed. This involves monitoring the trend of the vibration analysis data over time and identifying a problem after the point in time when this trend changes significantly.

However, for example, when producing a large number of screw compressors 10, the vibration analysis data varies from machine to machine (or loft), and this variation is actually large. Therefore, in order to eliminate such variations in data between machines, attention is being paid to performing the above-mentioned trend management using the fluctuation rate of the above-mentioned vibration analysis data. For example, the amplitude of the waveform data related to the vibration of one screw compressor 10 is 50
In the case of microns or more, this can be used as a criterion for determining rotor contact. However, some machines have an amplitude of 60 microns from the beginning even though it is normal. Therefore, in such a case, it is preferable to apply the fluctuation rate over time rather than the absolute value of the vibration analysis data.

However, when performing trend management using the fluctuation rate, the characteristics of the fluctuation rate for each machine may differ depending on the location of the abnormal order of the screw compressor 10. For example, there are universalities among machines regarding bearings, shaft seals, timing gears, etc.

Regarding the rotor, even when the above variation rate was applied, it was not possible to eliminate the variation between machines.

Therefore, even if the vibration analysis data from the screw compressor 10 was applied as input data to the neural network, it was not possible to properly diagnose the cause of the abnormality in the screw compressor 10 using this neural network. .

Therefore, one object of the present invention is to input vibration analysis data with little variation from loft to loft of the three-rotating machine into a neural network, thereby making it possible to diagnose the cause of abnormalities in the rotating machine with high precision. The objective is to provide an abnormality diagnosis device.

[Means to solve the problem]

The main means adopted by the present invention in order to achieve the above object is that a pair of rotors that engage with each other are rotatably supported in a bearing part in a casing, and fluid is compressed by the rotation of the rotors. Using Neural 7 Notework, a learning step is performed based on a backpropagation type learning algorithm in which input data is vibration analysis data from a rotating machine that performs The vibration analysis data for each cause of the abnormality is input to the neural network, and the connection weight of the neural network is changed in order to reduce the difference between the output data related to the cause of the abnormality at this time and the desired output data related to the cause of the abnormality. decided,
In the abnormality diagnosis device, which outputs output data obtained based on the determined connection weights as a cause of abnormality in the rotating machine when new vibration analysis data of the rotating machine is input to the neural network. The present invention is configured as an abnormality diagnosing device for a rotating machine using, as input vibration analysis data, data relating to a fluctuation rate based on the normal vibration analysis data of the rotating machine.

[Operation] According to the present invention, the fluctuation rate based on normal vibration analysis data of a 9-rotation machine is used as vibration analysis data input to a neural network in which a learning step is performed based on a backpropagation type learning algorithm. Since the data relating to this is used, variations in vibration analysis data for each loft of the rotating machine are eliminated.

(Example) Hereinafter, an example embodying the present invention will be explained with reference to the attached drawings to provide an understanding of the present invention. Fig. 1 shows a screw compressor according to an embodiment of the present invention. 2 is a conceptual diagram showing a bank propagation type neural network included in the abnormality diagnosing device, and FIG. 3 is a diagram showing the neurons and connections that make up the neural network. Conceptual diagram.

Fig. 4 is an explanatory diagram showing the process of enveloping the waveform data from the vibration sensor, and Fig. 5 is an explanatory diagram showing the waveform data from the vibration sensor and the probability density function for the amplitude direction of the waveform data at the same time. 6 are trend management graphs showing temporal changes in vibration analysis data analyzed by a waveform analyzer.

Note that the following examples are specific examples of the present invention.
It is not intended to limit the technical scope of the present invention.

Further, the same reference numerals are used for elements common to those of the screw compressor 10 shown in FIG. 7, and detailed explanation thereof will be omitted.

Abnormality diagnosis device 1 for screw compressor 10 according to the present embodiment
As shown in FIG. 1, the system uses a hack propagation type learning algorithm to learn the correspondence between certain input data from the screw compressor 10 and the cause of abnormality in the screw compressor 10 corresponding to this input data. Then, the connection weight of the connection unit 6 (FIG. 2) of the neural network is determined.

A conceptually three-layered neural network 2 outputs output data y (Figure 2) related to the cause of the abnormality based on the connection weights during diagnosis, and stores the connection weights determined during learning of the neural network 2. The memory 3 calculates the vibration-related data from the screw compressor 10 into data related to the rate of fluctuation based on the data when the screw compressor 10 is normal, and generates the neural network.
an arithmetic control unit 4 that inputs data to the sitwork 2, diagnoses the cause of abnormality in the screw compressor 10 based on output data from the neural network 2, and outputs the data to the screw compressor 10; The input/output port 12 is interposed between the screw compressor 10 and the input/output port 12 to transfer data.

As shown in FIGS. 2 and 3, the neurons 5 constituting the neural network 2 are connected to a connecting section 6 that transmits data signals and performs weighting processing on the data signals.

Neuron 5 in the upstream layer in the data transmission direction (arrow F)
The input section 7 inputs the input data from the connection section 6, and the weighting inputted to the input section 7 performs threshold processing on the sum value obtained by summing each subsequent input data. An arithmetic unit 8 which conceptually fires (outputs) when the value exceeds a threshold value set in advance for each neuron 5.
and an output section 9 that outputs a data signal to the neuron 5 in the downstream layer in the data transmission direction when the arithmetic section 8 fires.

At appropriate locations in the screw compressor 10.

The above casing 20 and bearing part a, female rotor 24.2
A magnetic vibration sensor 11 for measuring the shaft vibration of 5 is removably fixed.

The waveform data related to the vibration of the screw compressor IO detected by the vibration sensor 11 is input to the subsequent waveform analyzer 13. This waveform analyzer 13 is a general-purpose one, and performs predetermined waveform processing on the waveform data from the vibration sensor 11, and transfers it to the abnormality diagnosis device 1.
The vibration analysis data is converted into various types of vibration analysis data that can be used to diagnose the cause of an abnormality in the screw compressor 10, and is output to the abnormality diagnosis device 1.

According to this waveform analyzer 13, as shown in FIG. 4, for example, waveform data a having irregular vibrations centered around a high impact frequency f is inputted as waveform data from the vibration sensor 11. case.

Even if this waveform data a is expanded as it is as a frequency spectrum, it is not possible to obtain a spectrum peak with nine periods T. however.

To know the period T8 of such irregular vibrations.

Bearings 34, 35, 36, 3 of the above screw compressor IO
8.39. Shaft seal 28. 29.30. 31. This is important in detecting abnormalities in timing gears 32 and 33. Therefore, when absolute value processing is applied to the waveform data a with irregular vibrations as described above, the waveform data a is converted to waveform data b that has more periodicity. Further, when the waveform data b is subjected to envelope processing by passing through a low-pass filter, waveform data C having nine periods of vibration can be obtained. This waveform data C is waveform data in the time domain, and by subjecting it to Fourier transform, which is a linear reversible transform, a frequency spectrum in the frequency domain can be obtained. Then, the period T3 can be calculated from the frequency spectrum.

The abnormality diagnosis device 1 of this embodiment has two methods for analyzing waveform data related to vibrations of the screw compressor 10: one that analyzes the waveform data in the one-hour domain, and the other that analyzes the waveform data derived from the waveform data in the time domain. Some methods perform analysis based on the results of analysis based on waveform data in the frequency range.

Therefore, when waveform data a1 related to the time domain as shown in FIG. 5 appears, for example, regarding the vibration of the screw compressor 10.

For the former, the calculation result based on statistical processing of the amplitude x (t) of this waveform data a can be used.

In the figure, one curve b1 is the amplitude x of the waveform data a.
Indicates the probability density interval P (xl) for (t), in this case the waveform data a.

The vibration is a Gaussian irregular vibration, and this curve b1 has a normal distribution. In addition, the curve b2 shown by the − dotted chain line
is the amplitude x (t) when the vibration at that time is harmonic vibration
It shows the probability density function P (x) of .

If Gaussian irregular vibrations are applied to this system in advance and the probability density function P(2) of the normal distribution is obtained, then based on this probability density function P(x), the following basic quantities related to vibrations can be obtained: It is possible to calculate vibration analysis data such as (amplitude = x).

Standard deviation δ: Standard deviation related to amplitude X.

Average value: Ma = 5. : x P (x) d
x (3) Effective value = 0F577 for X (4) Skewness x2 is an index indicating the symmetry of the distribution.

Footosis: T' = s-: x' P(x)
dx...(6) Footsis x4 represents the degree of spread and sharpness of the 9 distribution, and is effective in diagnosing rolling bearings.

Crest factor: cF=p/xr+ms=・
cy) p is the peak value of amplitude X.

The effective value x ras is related to the sum of the power of all harmonics and is related to the magnitude of each harmonic.

Skewness: ”;13=5.:, x3P(x)d
x...(5) In addition, for the latter analysis method, the waveform data a
, is a method of performing frequency analysis based on vibration analysis data after Fourier transform.

For example, if amplitude-modulated time-domain waveform data appears, it is possible that misalignment or uneven contact exists in the timing gear bearing or bearing seal of the screw compressor 10. Therefore, when a frequency spectrum with displacement amplitude on the vertical axis and frequency on the horizontal axis is taken based on the above waveform data, side pads appear at both ends of the center frequency peak of the frequency spectrum.

Alternatively, when the waveform data is subjected to frequency modulation, there is a large pinch error in a gear device such as the timing gear.

In such a case, an infinite number of sidebands appear on both sides of the center frequency peak in the corresponding frequency spectrum. That is, it is also possible to detect an abnormal location in the screw compressor 10 based on the appearance position and number of side hands in the frequency spectrum.

Further, by performing frequency analysis on the waveform data after enclosing the waveform data in the time domain, it is also possible to detect the rotational frequency component, natural frequency, etc. of the 9-rotation device.

Therefore, the waveform analyzer 13 analyzes (1) to
The vibration analysis data related to equation (7), the vibration analysis data related to the side hand, etc. are manually input to the calculation control unit 4 as vibration analysis data xJ. And these vibration analysis data
J is subjected to averaging processing using the normal value as a reference.

1. For example, the screw compressor 10 of a certain lot (machine) is operated in a factory for an appropriate period (for example, one month after completion), the vibration analysis data XJ at that time is calculated, and the time is calculated as shown in Fig. 6. Normal vibration analysis data up to L1
The average value AJ of , is determined in advance and stored in the memory 3. Such averaging processing is performed in the screw compressor 10.
This is done for each loft and for each type of vibration analysis data XJ.

Then, each vibration analysis data xJ obtained during the above operation is applied to the following equation.

D, = (XJ/AJ)-1... (8) The fluctuation rate DJ of the vibration analysis data XJ based on each of the above average values AJ is obtained. This fluctuation rate DJ is.

The data is input as input data to the neuron 5 (FIG. 2) of the input layer of the neural network 2 of the abnormality diagnosis device 1. Such averaging processing is performed for each lot and for each type of vibration analysis data XJ. Therefore, the above fluctuation rate D
Since J is calculated based on the normal average value AJ of the vibration analysis data xJ for each lot, the data has no difference between lofts. That is, the value of the fluctuation rate DJ when the screw compressor 10 is normal is l for each lot. In addition,
The above-mentioned averaging process is not limited to applying the fluctuation rate D, but may also apply the fluctuation difference (XJ-A,) between the above-mentioned Xj and Aj.

Therefore, the abnormality diagnosis device 1 detects the screw compressor 1.
The procedure for diagnosing the abnormality of 0 will be explained below.

First, the learning step of the neural network 2 is executed. When the cause of the abnormality of the screw compressor 10 is clear 1 For example, when the cause of the abnormality is artificially set, the vibration analysis data As shown in FIG. 2, the time fluctuation rate DJ is input to the neuron 5 of the input layer of the neural network 2, and the cause of the abnormality or normal state at this time is set as teacher data. It should be noted that it is not necessary to use the fluctuation rate D of all types of vibration analysis data xJ, such as the effective value, skewness, and cutosis, as input data for the neural network 2. Only data related to characteristic vibration analysis data may be used.

Therefore, the fluctuation rate D1 input to the neuron 5 of the input layer
are sequentially calculated from the intermediate layer to the output layer in the data transmission direction (arrow F) based on the threshold set in the calculation unit 8 of the neuron 5 and the connection weight set in the connection unit 6, The output layer neuron 5 outputs the output data y. and.

The output data y at this time is compared with the teacher data set above, and the connection weight of the connection unit 6 is changed and set so that the output data y matches or approximates the teacher data. Such learning of connection weights is performed by sequentially backpropagating each layer toward the upstream side in the data transmission direction.

The connection weight determined by the connection unit 6L is used as an initial value during learning of the connection weight given to this connection unit 6 when primary learning human data is input.

Such connection weight learning is performed for multiple types of abnormal causes or normal states of the screw compressor 10 and the fluctuation rate DJ of the vibration analysis data XJ at that time, and the connection weights of each of the connection parts 6 are finally determined. and stored in the memory 3. In other words, the normal vibration analysis data X as described above
Information regarding the correspondence between the fluctuation rate with respect to the average value of J and the state of the screw compressor 10 at this time is obtained from each of the above-mentioned connecting portions 6.
In the end, it is conceptually distributed and stored in the connection weights of .

The steps from the sampling and averaging process of the vibration analysis data XJ to the learning of the neural network 2 are performed, for example, before the screw compressor 10 is shipped from the factory.

After the screw compressor 10 is installed at a predetermined location, for example, at a delivery destination, a test run thereof is performed.

Therefore, screw compression ii during the above test run.

Assuming that the vibration analysis data XJ is normal, the sampling and averaging process of the vibration analysis data
is stored in memory 3.

Subsequently, an abnormality diagnosis of the screw compressor 10 is performed using the learned neural network 2. The fluctuation rate DJ obtained from the vibration analysis data X from the screw compressor 10 and the normal average value A is input to the neural network 2. Therefore, the neural network 2 subjects the fluctuation rate D to calculation based on the connection weight of the connection unit 6 stored in the memory 3 and outputs output data y to the calculation control unit 4. The calculation control unit 4 determines the abnormality diagnosis result from the output data y and outputs it to the screw compressor 10.

As described above, the screw compressor 10 according to one embodiment
The abnormality diagnosis device 1 calculates the fluctuation rate based on the average value of normal vibration analysis data from the screw compressor lO every +7 tons and for each type of vibration analysis data. The variation in data between 100 lots of compressors is reduced. Therefore, by using the data with the above-mentioned fluctuation rate with small variation as input data for neural network learning and abnormality diagnosis, The condition can be diagnosed with high accuracy.

In addition, in the above-mentioned embodiment, it was stated that the fluctuation rate or fluctuation difference based on the normal average value of vibration analysis data can be used, but the above-mentioned neural network can also be used as an averaging process for such data. The threshold value of the arithmetic unit 8 in the neuron 5 of the second input layer can also be set as the average value in the normal state.

〔Effect of the invention〕

According to the present invention, vibration analysis data from a rotating machine that rotatably supports a pair of rotors meshing with each other in a bearing part in a casing and compresses fluid by rotation of the rotors is used as input data, and the rotating machine C. Using a neural network in which learning is performed based on a cupropagustan type learning algorithm, inputting the vibration analysis data for each abnormality cause into the neural network, In order to reduce the difference between the output data related to the 207th abnormality cause and the desired output data related to the abnormality cause, the connection weight of the neural network is changed and determined, and the neural network is In the abnormality diagnosis device that outputs output data obtained based on the determined connection weight as the cause of the abnormality of the rotating machine when vibration analysis data is input, the vibration analysis data input to the neural network is An abnormality diagnosing device for a rotating machine is provided, which uses data related to a fluctuation rate based on vibration analysis data of the rotating machine during normal operation. Thereby, it is possible to input into the neural network data related to the fluctuation rate of vibration analysis data with little variation from loft to loft (or machine) to loft (or machine) of the one-rotation machine. As a result, the cause of the abnormality in the rotating machine can be diagnosed with high accuracy without lot-to-lot variation.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an abnormality diagnosis device for a screw compressor according to an embodiment of the present invention, and FIG. 2 is a conceptual diagram showing a back-propagation type neural network included in the abnormality diagnosis device. Part 3 is a conceptual diagram showing the neurons and connections that make up the neural network, Part 4 is an explanatory diagram showing the process of applying envelope processing to waveform data from a vibration sensor, and Fig. 5 is a conceptual diagram showing the neurons and connections that make up the neural network. An explanatory diagram that simultaneously shows the waveform data from the vibration sensor and the probability density function for the amplitude direction of the waveform data, Fig. 6 is a trend management graph showing the time change of vibration analysis data analyzed by the waveform analyzer, and Fig. 7 is the screw pressure 1ife!, which is an example of the background of the present invention! FIG. [Explanation of symbols] 1...Abnormality diagnostic device 2...Neural network 3...Memory 4...Arithmetic control unit 10...Screw compressor 20...Casing 24...Male rotor 25 ...Female rotor 26.27...Rotor shaft

Claims (1)

[Claims] (1) A pair of rotors that mesh with each other are rotatably supported on bearings in the casing, and vibration analysis data from a rotating machine that compresses fluid by the rotation of the rotors is used as input data, and the causes of abnormalities in the rotating machine are determined. Using a neural network in which a learning step is performed based on a backpropagation type learning algorithm that uses data related to as output data, the vibration analysis data for each cause of the abnormality is input to the neural network, and the vibration analysis data for each cause of the abnormality at this time is input to the neural network. When the connection weight of the neural network is changed and determined in order to reduce the difference between the output data and the desired output data regarding the cause of the abnormality, and new vibration analysis data of the rotating machine is input to the neural network. In an abnormality diagnosis device that outputs output data obtained based on the determined connection weights as the cause of the abnormality of the rotating machine, vibration analysis data of the rotating machine during normal operation is inputted to the neural network as vibration analysis data. An abnormality diagnosis device for a rotating machine, characterized in that it uses data related to a rate of fluctuation based on .