Abnormality diagnosis device and abnormality diagnosis method

Abstract

To reduce calculation cost and recording cost of re-learning.SOLUTION: A plurality of feature amounts are calculated with respect to waveform data divided into a plurality of parts, for generating a feature amount matrix, then a feature amount average vector and a feature amount standard deviation vector for every kind of the feature amount are generated from the feature amount matrix, then, the generated feature amount average vector and the feature amount standard deviation vector are compared with existing learning data, for calculating an abnormality degree of a current diagnosis object for diagnosing an abnormality. Then, on the basis of the generated feature amount average vector and the feature amount standard deviation vector, a division number of the waveform data, the feature amount average vector and the feature amount standard deviation vector of the existing learning data, and a total sample number, a feature amount average vector and a feature amount standard deviation vector and a total sample number of a new learning object are calculated, and a result of update of the calculation is recorded.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to an abnormality diagnostic device and an abnormality diagnostic method for a rotating machine, a load device, or the like. [Background technology]

[0002] 2. Description of the Related Art Equipment used in fields such as manufacturing and social infrastructure is often operated continuously for long periods of time, and therefore requires periodic inspection or constant monitoring for abnormality diagnosis. Generally, it is difficult to temporarily suspend the operation of such equipment, and there is a demand for abnormality diagnosis that can be performed while the equipment is in operation. An example of such equipment is a rotating machine equipped with at least one of a sliding bearing, a rolling bearing, and a magnetic bearing. Conventionally, a widely used technique for detecting abnormalities in rotating machines during operation involves processing vibration waveform data collected by a vibration sensor and detecting signs of abnormality.

[0003] Patent Document 1, which is an example of the prior art, discloses an equipment abnormality diagnosis device and an abnormality diagnosis method that can accurately identify the cause of an abnormality in equipment during operation. In Patent Document 1, a learning phase and a diagnosis phase are separated, learning of a normal state is performed in advance, and any deviation from the normal state is regarded as an abnormal state. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2020-8337 A Summary of the Invention [Problem to be solved by the invention]

[0005] However, in such an abnormality diagnosis device, the normal state of the diagnosis target changes in response to changes in the operating environment, and so it is necessary to successively re-learn new normal states to update the normal state. The above-mentioned conventional techniques have a problem in that, in order to perform such re-learning, it is necessary to keep records of the waveform data or all of the feature amount data used for learning.

[0006] The present invention has been made in view of the above, and has an object to reduce the load of relearning by reducing the calculation cost and recording cost for relearning. [Means for solving the problem]

[0007] One aspect of the present invention for solving the above-mentioned problems and achieving the object includes a data division unit that divides waveform data acquired from a diagnosis target into a plurality of waveform data by frequency or time, a feature calculation unit that calculates a plurality of feature amounts for each of the plurality of waveform data and generates a feature matrix by the calculated feature amounts, a feature statistics unit that calculates an average and a standard deviation in the row direction for the feature matrix and generates a feature average vector and a feature standard deviation vector for each type of feature amount, and a feature average vector and a feature standard deviation vector generated by the feature statistics unit and a feature average vector and a feature standard deviation vector for existing learning data. an abnormality diagnosis unit that calculates a degree of abnormality of the current diagnosis object by comparing the degree of abnormality with the current learning data and diagnoses whether the diagnosis object is normal or abnormal based on the degree of abnormality; a learning update unit that calculates a feature average vector and a feature standard deviation vector and a total number of samples of a new learning object based on the feature average vector and the feature standard deviation vector generated by the feature statistics unit, the number of divisions of the waveform data, the feature average vector and the feature standard deviation vector of the existing learning data, and the total number of samples; and a learning information recording unit that records the update results in the learning update unit.

[0008] In the abnormality diagnosis device having the above configuration, it is preferable that the data dividing section divides the waveform data into a plurality of waveform data while excluding peaks of the waveform data.

[0009] Alternatively, one aspect of the present invention for solving the above-mentioned problems and achieving the object is an anomaly diagnosis method including: dividing waveform data acquired from a diagnosis target into a plurality of waveform data by frequency or time; calculating a plurality of feature amounts for each of the plurality of waveform data and generating a feature amount matrix by the calculated feature amounts; taking an average and a standard deviation in the row direction for the feature amount matrix to generate a feature amount average vector and a feature amount standard deviation vector for each type of feature amount; comparing the generated feature amount average vector and the feature amount standard deviation vector with a feature amount average vector and a feature amount standard deviation vector of existing learning data to calculate a degree of abnormality of the current diagnosis target, and diagnosing whether the diagnosis target is normal or abnormal based on the degree of abnormality; calculating a feature amount average vector and a feature amount standard deviation vector of a new learning target, and a total number of samples, based on the generated feature amount average vector and the feature amount standard deviation vector, the number of divisions of the waveform data, the feature amount average vector and the feature amount standard deviation vector of the existing learning data, and a total number of samples; and recording an update result obtained by the calculation.

[0010] In the abnormality diagnosis method configured as described above, it is preferable that when the waveform data is divided into the plurality of waveform data, peaks of the waveform data are excluded before the waveform data is divided into the plurality of waveform data. Effect of the Invention

[0011] According to the present invention, the computation cost and recording cost for re-learning are reduced, thereby reducing the load of re-learning. [Brief description of the drawings]

[0012] [Figure 1] FIG. 1 is a diagram showing an abnormality diagnosis device according to an embodiment and a peripheral configuration including an object to be diagnosed by this abnormality diagnosis device. [Diagram 2] FIG. 2 is a functional block diagram showing the configuration of the abnormality diagnosis device according to the embodiment. [Diagram 3]FIG. 3 is a flowchart showing an abnormality diagnosis method when the abnormality diagnosis device according to the embodiment collects new data to perform abnormality diagnosis and learning update. [Figure 4] FIG. 4 is a diagram showing an example in which a frequency filter in the second frequency filter range is applied to a typical example of frequency characteristics of vibration data measured in a normal state by an acceleration sensor in a rotating machine to be diagnosed. [Diagram 5] FIG. 5 is a diagram showing an example in which a frequency filter in a first frequency filter range is applied to a typical example of frequency characteristics of vibration data measured in a normal state by an acceleration sensor in a rotating machine to be diagnosed. [Figure 6] FIG. 6 is a conceptual diagram showing an image of a shock waveform produced by a damped sine wave. [Figure 7] FIG. 7 is a graph showing a comparison of the degree of abnormality in the diagnosis of a simulated bearing abnormality using a 5 kHz damped sine wave. [Figure 8] FIG. 8 is a graph showing a comparison of the degree of abnormality in the diagnosis of a simulated bearing abnormality using an 8 kHz damped sine wave. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] Hereinafter, an embodiment of the present invention will be described with reference to the accompanying drawings. However, the present invention should not be construed as being limited to the following embodiments.

[0014] (Embodiment) FIG. 1 is a diagram showing an abnormality diagnostic device 1 according to this embodiment and a peripheral configuration including an object to be diagnosed by the abnormality diagnostic device 1. As shown in FIG. The abnormality diagnostic device 1 shown in FIG. The sensors 4 attached to the rotating machine 2 and the load device 3 detect the states of the rotating machine 2 and the load device 3 . The detection value of the sensor 4 is acquired by a data measuring instrument 5 . The data measuring instrument 5 collects measurement data, which are values ​​detected by the sensor 4 , and sends the data to the abnormality diagnosis device 1 . The abnormality diagnostic device 1 uses data collected by the data measuring instrument 5 to learn normal conditions and perform abnormality diagnosis.

[0015] In FIG. 1, a vibration sensor is shown as the sensor 4 and installed near a bearing of the rotating machine 2 or the load device 3 to be diagnosed, but the present invention is not limited to this. Examples of the sensor 4 include a vibration sensor, an acoustic sensor that measures the sound of the diagnosis target, and a current sensor that measures the power supply current or leakage current of the diagnosis target.

[0016] In FIG. 1, the result of an abnormality diagnosis by an abnormality diagnosis device 1 is displayed on a display device 6 directly connected to the abnormality diagnosis device 1 . However, the present invention is not limited to this, and the abnormality diagnosis results may be sent to a SCADA (Supervisory Control And Data Acquisition) connected via a network and displayed on a display device within the facility, or sent to the cloud and displayed on a display device outside the facility.

[0017] FIG. 2 is a functional block diagram showing the configuration of the abnormality diagnosis device 1 according to this embodiment. FIG. 3 is a flowchart showing an abnormality diagnosis method when the abnormality diagnosis device 1 according to this embodiment collects new data and performs abnormality diagnosis and learning update. The abnormality diagnosis device 1 shown in FIG. 2 includes a data collection unit 11, a data division unit 12, a feature calculation unit 13, a feature statistics unit 14, an abnormality diagnosis unit 15, a diagnosis result transmission unit 16, a learning update unit 17, and a learning information recording unit 18.

[0018] The data collection section 11 collects and outputs data (waveform data) from the data measurement instrument 5. The waveform data to be collected may be, for example, data sampled at 51200 Hz (51.2 kHz) for 5 seconds each time. When the diagnosis target has a bearing abnormality or the like, vibrations up to about 10 kHz are present, so sampling at 20 kHz or higher is preferable. However, the present invention is not limited to this, and the sampling frequency may be appropriately determined depending on the type of abnormality to be diagnosed. The waveform data is preferably collected while the rotating machine 2 or the load device 3 to be diagnosed is in operation. Furthermore, limiting the operating conditions can improve diagnostic accuracy.

[0019] The data dividing unit 12 divides the waveform data (waveform data acquired from a diagnostic subject) collected and output by the data collecting unit 11 into a plurality of waveform data based on frequency or time, and outputs the data (S1).

[0020] When the data dividing section 12 divides the waveform data in the time direction, that is, by time, for example, the data dividing section 12 divides the entire waveform data in chronological order. Here, for example, 5 seconds of data is divided into five waveform data segments of 0 to 1 second, 1 to 2 seconds, 2 to 3 seconds, 3 to 4 seconds, and 4 to 5 seconds. Alternatively, when dividing the data, some of the data may overlap; for example, 5 seconds of data may be divided into nine pieces of waveform data: 0 to 1 second, 0.5 to 1.5 seconds, 1 to 2 seconds, 1.5 to 2.5 seconds, 2 to 3 seconds, 2.5 to 3.5 seconds, 3 to 4 seconds, 3.5 to 4.5 seconds, and 4 to 5 seconds.

[0021] Alternatively, when the data division unit 12 divides the waveform data in the frequency direction, i.e., by frequency, for example, a frequency filter is applied to the waveform data to divide it into three waveform data: low frequency, medium frequency, and high frequency. An example of a frequency filter is a third-order Chebyshev 1 filter (ripple is 1 dB). Here, it is preferable that there are no peaks in the waveform data in the low, medium and high frequency regions. For example, it is preferable that a low frequency is greater than or equal to 5 Hz and less than 250 Hz, a medium frequency is greater than or equal to 300 and less than 3000 Hz, and a high frequency is greater than or equal to 5000 Hz (first frequency filter range). In this way, excluding peaks of the waveform data from the frequency domain improves the judgment sensitivity.

[0022] FIG. 4 is a diagram showing an example in which a frequency filter in the second frequency filter range is applied to a typical example of frequency characteristics of vibration data measured in a normal state by an acceleration sensor in the rotating machine 2 to be diagnosed. The peaks in the vibration data in FIG. 4 are present in the vicinity of 20 to 300 Hz and above 3000 Hz.

[0023] FIG. 5 is a diagram showing an example in which a frequency filter in the first frequency filter range is applied to a typical example of frequency characteristics of vibration data measured in a normal state by an acceleration sensor in a rotating machine 2 to be diagnosed. In FIG. 5, in the medium frequency filter range, the frequency components of the waveform data in a normal state are low, and the peaks present near 300 Hz and near 3000 Hz are removed by the frequency filter.

[0024] However, the present invention is not limited to this, and a low frequency may be 5 Hz or more and less than 1000 Hz, a medium frequency may be 1000 Hz or more and less than 5000 Hz, and a high frequency may be 5000 Hz or more (second frequency filter range).

[0025] The feature amount calculation section 13 calculates a plurality of feature amounts for each of the plurality of waveform data obtained by the data division section 12, and generates and outputs a feature amount matrix (S2). In calculating the feature amount, the same feature amount may be calculated for the waveform data divided in time order, and different feature amounts may be calculated for the waveform data divided by frequency. In the feature matrix, feature amounts for each type are arranged in the column direction, and feature amounts divided by time are arranged in the row direction. Here, feature quantities with different frequency conditions are treated as different types.

[0026] The feature quantity statistics unit 14 calculates the average and standard deviation row-wise for the feature quantity matrix output from the feature quantity calculation unit 13, and generates and outputs a feature quantity average vector and a feature quantity standard deviation vector for each type of feature quantity (S3).

[0027] The abnormality diagnosis unit 15 compares the feature average vector and the feature standard deviation vector generated by the feature statistics unit 14 with the feature average vector and the feature standard deviation vector of existing learning data, and calculates the degree of abnormality of the waveform data of the current diagnosis target from the difference between the two (S4), and diagnoses whether the diagnosis target is normal or abnormal based on this degree of abnormality, and outputs the abnormality diagnosis result (S5). If the calculated abnormality degree is less than a preset threshold, it is determined to be normal, and if it is equal to or greater than the preset threshold, it is determined to be abnormal. The method for calculating the degree of abnormality is disclosed in Patent Document 1.

[0028] The diagnosis result transmission section 16 outputs the abnormality diagnosis result output by the abnormality diagnosis section 15 to the display device 6 . The diagnosis result transmission unit 16 may output the waveform data measured by the sensor 4 itself or a plurality of waveform data divided by the data division unit 12 together with the abnormality diagnosis result.

[0029] The learning update unit 17 calculates the feature average vector and feature standard deviation vector and the total number of samples of a new learning object based on the feature average vector and feature standard deviation vector generated in the feature statistics unit 14, the number of divisions of the waveform data, the existing feature average vector and feature standard deviation vector of the learning object, and the total number of samples, and outputs the update result to the learning information recording unit 18 (S6). Here, it is preferable that the learning update in the learning update section 17 is performed only when the abnormality diagnosis result in the abnormality diagnosis section 15 is normal.

[0030] The learning information recording unit 18 records the update result in the learning update unit 17 as learning information, which is learning data (S7).

[0031] In the abnormality diagnostic device 1 according to this embodiment, in addition to the effective value, the feature quantities to be calculated can be those described as diagnostic parameters in Patent Document 1. The diagnostic parameters described in Patent Document 1 are shown in Table 1 below. The formulas for calculating the parameters shown in Table 1 are disclosed in Patent Document 1.

[0032] [Table 1]

[0033] For example, at low frequencies, the effective value, p4 (crest factor), p12 (ratio of zero-value crossing frequency to absolute maximum frequency), p13 (ratio of zero-value crossing frequency to valley frequency) and p14 (ratio of absolute maximum frequency to valley frequency) are used, while at medium and high frequencies, the effective value, p1 (rate of fluctuation), p15 (average square root of the waveform) and p20 (kurtosis in the frequency domain) are used. The number of features used is 5+4+4=13.

[0034] Next, the calculation formula for learning update executed by the learning update unit 17 will be described. Using data on the learning object and data on the waveform data to be newly added as a learning object, the newly recorded learning object data is expressed by the following formulas (1) to (3). Here, the data for the learning target is the feature mean m L , feature standard deviation σ L 2 and the total number of samples is l. The waveform data newly added as the learning target is the feature average m S , feature standard deviation σ S 2 and the number of time divisions is s. The newly recorded learning data is N , feature standard deviation σ N 2 and the total number of samples is n.

[0035]

number

[0036]

number

[0037]

number

[0038] Here, the derivation of the above formulas (1) to (3) will be explained below. First, the features of the existing learning object are the sequence X=x1,…,x l The features of the additional learning objects are represented as a sequence Y=y1,…,y s Then, the updated features of the learning object are expressed as the sequence Z=x1,…,x l ,y1,…,y s It is expressed as: Then, the feature average m L ,m S ,m N Equation (2) is obtained from

[0039]

number

[0040] The feature standard deviation σ L 2 Therefore, the feature standard deviation σ N 2 is obtained.

[0041]

number

[0042] Although the above-mentioned formula is for each feature, similar calculations can be performed for each of a plurality of feature amounts, and the formula can be similarly extended even when the target is a feature amount vector.

[0043] As described above, according to this embodiment, waveform data obtained by a sensor is divided, and from the multiple waveform data obtained by division, multiple types of feature quantities are calculated for each division in the time direction, and the average and standard deviation are calculated from multiple samples for each type of feature quantity, thereby obtaining an approximate distribution of each feature quantity. Here, the number of samples for calculating the average and standard deviation of the feature amount in learning is (the number of waveform data used for learning)×(the number of divisions in the time direction). Then, the distribution of the mean and standard deviation obtained by learning for each feature amount is compared with the distribution of the mean and standard deviation obtained by diagnosis, and the degree of anomaly is calculated based on the difference between the two. Furthermore, if the calculated abnormality degree is less than a preset threshold, it is determined to be normal, and if it is equal to or greater than the preset threshold, it is determined to be abnormal.

[0044] According to this embodiment, by recording the feature mean vector, feature standard deviation vector, and total sample number of the learning target, it is possible to update the learning of the feature mean vector, feature standard deviation vector, and total sample number including additional waveform data. This eliminates the need to record the waveform data itself that is the subject of learning, making it possible to reduce the computational costs and recording costs of learning.

[0045] It should be noted that the present invention is not limited to the above-described embodiment, but includes various modified examples in which components are added, deleted, or converted from the above-described configuration. EXAMPLES

[0046] In this embodiment, the effectiveness of the frequency filter provided in the data division unit 12 is explained, which has the above-mentioned first frequency filter range, i.e., low frequencies between 5 Hz and 250 Hz, medium frequencies between 300 and 3000 Hz, and high frequencies above 5000 Hz.

[0047] In this embodiment, 48 sets of vibration data measured under normal conditions by the acceleration sensor, which is the sensor 4, of the rotating machine 2 are used as learning data, and artificial abnormal waveform data in which repeated shock waveforms of damped sine waves simulating bearing abnormalities are superimposed at various magnitudes on the vibration data of the rotating machine 2 in a similar normal state is diagnosed. The degree of abnormality when a frequency filter in the first frequency filter range is used is compared with that when a frequency filter in the second frequency filter range is used.

[0048] First, FIG. 6 is a conceptual diagram showing an image of a shock waveform produced by a damped sine wave. The damped sine wave indicates a waveform obtained by multiplying an exponential function by a repeating sine wave of a damped sawtooth wave. In Figure 6, the peak decays to half for each cycle of the sine wave, and a new shock wave is created every five cycles, so even in the most decaying part, the amplitude remains a few percent of the peak. However, in an actual shock wave, the sine wave frequency is several kHz, the interval between shocks, i.e., the pass frequency, is several tens of Hz, and the shock wave decays over about 100 periods. Therefore, the amplitude is generally zero except immediately after the impact. The magnitude of the damped sine wave to be superimposed is represented by the ratio of the peak value of the damped sine wave to the effective value of the original waveform.

[0049] FIG. 7 is a graph showing the degree of abnormality in the diagnosis of a simulated bearing abnormality using a 5 kHz damped sine wave. FIG. 8 is a graph showing the degree of abnormality in the diagnosis of a simulated bearing abnormality using an 8 kHz damped sine wave. In Figures 7 and 8, the sine wave is 5 kHz or 8 kHz, the damping rate per sine wave period is 0.5, and the impact interval is 40 Hz. 7 and 8, a state with an abnormality degree of less than 20 is set as normal, and a state with an abnormality degree of 20 or more is set as abnormal.

[0050] Regarding the simulated bearing abnormality caused by a 5 kHz damped sine wave, the frequency filter in the second frequency filter range shown in Figure 7(A) determined that an abnormality occurred when the superimposed magnitude of the damped sine wave was approximately 0.8 or more, whereas the frequency filter in the first frequency filter range shown in Figure 7(B) determined that an abnormality occurred when the superimposed magnitude of the damped sine wave was approximately 0.7 or more. In other words, the determination sensitivity in the first frequency filter range is improved by approximately 0.1.

[0051] Regarding the simulated bearing abnormality caused by an 8 kHz damped sine wave, the frequency filter in the second frequency filter range shown in Figure 8(A) judged an abnormality when the superposition magnitude of the damped sine wave was approximately 1.95 or more, whereas the frequency filter in the first frequency filter range shown in Figure 8(B) judged an abnormality when it was approximately 1.45 or more. In other words, the judgment sensitivity in the first frequency filter range is improved by approximately 0.5.

[0052] As described in this embodiment, it was confirmed that the determination sensitivity was improved by using a frequency filter in the first frequency filter range for both a 5 kHz attenuated sine wave and an 8 kHz attenuated sine wave. [Explanation of symbols]

[0053] 1. Abnormality diagnosis device 11 Data Collection Section 12 Data division section 13 Feature Calculation Unit 14 Feature Statistics Section 15 Abnormality diagnosis section 16 Diagnostic result transmission section 17 Learning and Updating Department 18 Learning Information Record Section 2 Rotating Machines 3 Load device 4 Sensors 5 Data measurement instruments 6 Display device

Claims

1. a data division unit that divides the waveform data acquired from the diagnosis target into a plurality of waveform data based on frequency or time; a feature calculation unit that calculates a plurality of types of feature values ​​including an effective value of the waveform for each of the plurality of waveform data, and generates a feature matrix in which the feature values ​​for each type are arranged in a column direction and the time is divided in a row direction; a feature quantity statistics unit that calculates an average and a standard deviation in the row direction of the feature quantity matrix to generate a feature quantity average vector and a feature quantity standard deviation vector for each type of the feature quantity; an anomaly diagnosis unit that calculates an anomaly level of the current diagnosis object by comparing the feature amount average vector and the feature amount standard deviation vector generated by the feature amount statistics unit with the feature amount average vector and the feature amount standard deviation vector of existing learning data, and diagnoses whether the diagnosis object is normal or abnormal based on the anomaly level; a learning update unit that calculates a feature quantity average vector, a feature quantity standard deviation vector, and a total number of samples of a new learning object based on the feature quantity average vector and the feature quantity standard deviation vector generated by the feature quantity statistics unit, the division number of the waveform data, the feature quantity average vector and the feature quantity standard deviation vector of the existing learning data, and the total number of samples; a learning information recording unit that records an update result in the learning update unit.

2. The abnormality diagnosis device according to claim 1 , wherein the data dividing unit divides the waveform data into a plurality of waveform data while excluding peaks of the waveform data.

3. Dividing the waveform data acquired from the diagnostic subject into a plurality of waveform data based on frequency or time; calculating a plurality of types of feature quantities including an effective value of the waveform for each of the plurality of waveform data, and generating a feature quantity matrix in which the feature quantities for each type are arranged in a column direction and the time is divided in a row direction; taking an average and a standard deviation in the row direction of the feature matrix to generate a feature average vector and a feature standard deviation vector for each type of feature; calculating an abnormality degree of the current diagnosis object by comparing the generated feature quantity average vector and the feature quantity standard deviation vector with the feature quantity average vector and the feature quantity standard deviation vector of existing learning data, and diagnosing whether the diagnosis object is normal or abnormal based on the abnormality degree; calculating a feature quantity average vector, a feature quantity standard deviation vector, and a total number of samples of a new learning object based on the generated feature quantity average vector and the feature quantity standard deviation vector, the division number of the waveform data, the feature quantity average vector and the feature quantity standard deviation vector of the existing learning data, and the total number of samples; and recording an updated result of said calculation.

4. The abnormality diagnosis method according to claim 3 , wherein the waveform data is divided into a plurality of waveform data while excluding peaks of the waveform data.

Images