Abnormality diagnostic device, abnormality diagnostic method, and program

Abstract

To diagnose the presence of an abnormality from an increase in stationary noise.SOLUTION: An abnormality diagnostic device comprises: an acquisition unit that acquires measurement data of vibration; a filtering unit that executes, on the measurement data, filtering using at least one of a Wiener filter, a Kalman filter, a regularized least squares method, an LMS adaptive filter, and a normalized LMS adaptive filter; a residual calculation unit that calculates a residual by applying a coefficient of an autoregressive model to the measurement data passing through the filtering unit; and a determination unit that determines the presence or absence of an abnormality on the basis of the residual.SELECTED DRAWING: Figure 4

Description

[Technical field]

[0001] The present invention relates to an abnormality diagnosis device, an abnormality diagnosis method, and a program. [Background technology]

[0002] 2. Description of the Related Art Abnormality diagnosis devices for detecting abnormalities in bearings and cages are known. For example, Patent Document 1 discloses a technique for detecting flaws in the inner ring or cage of a bearing from frequency characteristics using an autoregressive model. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2018-163135 A Summary of the Invention [Problem to be solved by the invention]

[0004] One of the abnormalities that requires diagnosis for bearings and the like is deterioration of the lubricant, but from the viewpoint of maintenance, even when the lubricant is deteriorated, it must be diagnosed as an abnormality before damage or seizure occurs in the bearing or cage. When damage occurs, a periodic abnormal sound occurs with the rotation of the shaft, which can be detected by the technology of Patent Document 1. However, before damage occurs, it is thought that only stationary noise (stationary noise) showing a Gaussian distribution increases with the deterioration of the lubricant, and no periodic abnormal sound occurs, and therefore it cannot be detected by the technology of Patent Document 1.

[0005] Therefore, an object of the present invention is to diagnose an abnormality based on an increase in stationary noise. [Means for solving the problem]

[0006] In order to solve the above problems, one aspect of an abnormality diagnosis device according to the present invention includes an acquisition unit that acquires vibration measurement data, a filter unit that filters the measurement data with at least one of a Wiener filter, a Kalman filter, a regularized least squares method, an LMS (Least Mean Square) adaptive filter, and a normalized LMS adaptive filter, a residual calculation unit that calculates a residual by applying coefficients of an autoregressive model to the measurement data that has passed through the filter unit, and a determination unit that determines whether the data is normal or abnormal based on the residual.

[0007] According to such an abnormality diagnosis device, the filter section applies so-called optimal filtering to the measurement data to remove wideband noise, and stationary noise components are extracted as residuals with the autoregressive model. Therefore, even if an abnormality occurs before scratches occur in the bearings or cage, such as deterioration of the lubricant, an increase in stationary noise showing a Gaussian distribution can be detected as an increase in the residuals, making it possible to diagnose the presence or absence of an abnormality.

[0008] In the above-mentioned abnormality diagnosis device, it is preferable to further include a pre-filter section that performs filtering to limit a frequency band of the measurement data before the filter section. By limiting the frequency band in the pre-filter section, the effects of unnecessary mechanical vibrations and the like are eliminated, improving the accuracy of abnormality diagnosis based on the residual.

[0009] In addition, it is also preferable that the above-mentioned abnormality diagnosis device further includes a model calculation unit that calculates the above-mentioned autoregressive model using measurement data of vibration under normal conditions as training data. As the autoregressive model, a general-purpose autoregressive model may be applicable, but a dedicated autoregressive model specialized for the operating environment may be required. By using an abnormality diagnosis device including a model calculation unit, a dedicated autoregressive model can be calculated, expanding the applicable operating environments.

[0010] In the abnormality diagnosis device, it is preferable that the model calculation unit calculates the autoregressive model using the training data that has passed through the filter unit. In the calculation of the autoregressive model, a more accurate autoregressive model can be obtained by removing wideband noise through optimal filtering by the filter unit.

[0011] In the above-mentioned anomaly diagnosis device, it is preferable that the model calculation unit calculates the autoregressive model using a least squares method. Since the training data that has passed through the filter unit has noise removed, an appropriate autoregressive model can be calculated by the least squares method.

[0012] In order to solve the above-mentioned problems, one aspect of an abnormality diagnosis method according to the present invention includes an obtaining step of obtaining vibration measurement data, a filtering step of filtering the measurement data with at least one of a Wiener filter, a Kalman filter, a regularized least squares method, an LMS adaptive filter, and a normalized LMS adaptive filter, a residual calculation step of calculating a residual by applying coefficients of an autoregressive model to the measurement data that has been subjected to the filtering step, and a determination step of determining whether the measurement data is normal or abnormal based on the residual.

[0013] According to this anomaly diagnosis method, broadband noise is removed in the filter step, so that an increase in stationary noise exhibiting a Gaussian distribution can be detected as an increase in the residual with the autoregressive model, making it possible to diagnose the presence or absence of an anomaly. In order to solve the above problems, one aspect of the program according to the present invention causes an information processing device to operate as each component of the above-mentioned abnormality diagnosis device. With such a program, the abnormality diagnosis device can be easily realized by the information processing device. Effect of the Invention

[0014] According to the present invention, an abnormality can be diagnosed based on an increase in stationary noise. [Brief description of the drawings]

[0015] [Figure 1] FIG. 2 is a bottom view showing a schematic diagram of an example of a vehicle. [Diagram 2] FIG. 1 is a side view illustrating an example of a vehicle. [Diagram 3] FIG. 1 illustrates an example of a hardware configuration of an abnormality diagnosis device. [Figure 4] FIG. 2 is a diagram illustrating a functional configuration of the abnormality diagnosis device. [Diagram 5] 4 is a flowchart showing an abnormality diagnosis process performed by the abnormality diagnosis device. [Figure 6] 4 is a graph showing an example of frequency characteristics of a response of an acceleration sensor. [Figure 7] 13 is a graph showing an example of residuals in a comparative example in which the optimal filter is omitted. [Figure 8] 11 is a graph showing an example of a residual in the present embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0016] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. However, in order to avoid the following description becoming unnecessarily redundant and to facilitate understanding by those skilled in the art, more detailed description than necessary may be omitted. For example, detailed description of already well-known matters and duplicated description of substantially the same configuration may be omitted. In addition, elements shown in the previously described figures may be appropriately referenced in the description of the later figures.

[0017] 1 and 2 are diagrams showing an example of a vehicle equipped with an embodiment of an abnormality diagnosis device of the present invention, Fig. 1 being a bottom view and Fig. 2 being a side view. 1 and 2 show a railway vehicle 100 as an example. The vehicle 100 has a structure in which one body 110 is supported by two bogies 111, and each bogie 111 is provided with two axles 102. A wheel 101 is fixed to each end of each axle 102, so that one vehicle 100 is provided with a total of eight wheels 101.

[0018] Axles 102 protrude outward from each wheel 101 and are rotatably supported by axle boxes 103. Each axle box 103 is equipped with an acceleration sensor 104, which measures the acceleration of vibration, for example in the vertical direction, and outputs a measurement signal. The measurement signal obtained by acceleration sensor 104 is input to an abnormality diagnosis device 105 provided in main body 110.

[0019] A bearing (not shown) is installed in each axle box 103, and a lubricant such as grease is applied (applied or injected) to the axle box 103. For maintenance purposes, it is desirable to replace the lubricant that has deteriorated to a certain extent, and an abnormality diagnosis device 105 diagnoses the deterioration of the lubricant as an abnormality based on the measurement signal obtained by the acceleration sensor 104.

[0020] The abnormality diagnosis device 105 is realized by, for example, installing and executing an embodiment of the program of the present invention in an information processing device such as an embedded computer. FIG. 3 is a diagram showing an example of a hardware configuration of the abnormality diagnostic device 105. As shown in FIG. The abnormality diagnosis device 105 includes an information processing unit 201 and a sensor data acquisition unit 202. The information processing unit 201 includes a CPU (Central Processing Unit) 211, a ROM (Read Only Memory) 212, a RAM (Random Access Memory) 213, and a DMA (Direct Memory Access) controller 214.

[0021] The information processing unit 201 executes abnormality diagnosis processing based on a program. The sensor data acquisition unit 202 is connected to the information processing unit 201 via, for example, a high-speed LAN line, samples the measurement signal input from the acceleration sensor 104 , acquires it as digital measurement data, and inputs it to the information processing unit 201 .

[0022] The CPU 211 executes an embodiment of the program of the present invention to process the measurement data and diagnose the deterioration of the lubricant. The ROM 212 stores the programs executed by the CPU 211 . The RAM 213 stores the above-mentioned measurement data and the like.

[0023] The DMA controller 214 stores the measurement data input from the sensor data acquisition unit 202 directly in the RAM 213 without passing through the CPU 211 . The information processing unit 201 may include a display unit that displays the diagnosis result by the information processing unit 201, and may include a communication unit that communicates the diagnosis result with an external computer. Furthermore, the abnormality diagnosis device 105 may be realized by dedicated hardware.

[0024] FIG. 4 is a diagram showing the functional configuration of the abnormality diagnostic device 105. The abnormality diagnostic device 105 includes a pre-filter unit 301 , an optimum filter unit 302 , a model identification unit 303 , a model inverse filter unit 304 , and a diagnostic unit 305 . 4 shows the functional configuration of the abnormality diagnosis device 105, and also shows the program configuration in one embodiment of the program of the present invention. The information processing device operates as each component of the abnormality diagnosis device 105 by the program having the program configuration shown in FIG.

[0025] The pre-filter unit 301 corresponds to an example of a combination of the acquisition unit and the pre-filter unit of the present invention, and acquires measurement data and performs filtering on the measurement data to limit the frequency band. The optimal filter unit 302 corresponds to an example of the filter unit referred to in the present invention, and removes wideband noise by performing optimal filtering on the measurement data using, for example, the regularized least squares method. Here, "optimal" means that a filter that is recognized as a general solution for optimization in filtering is used. In other words, it is sufficient to perform filtering using at least one of the Wiener filter, Kalman filter, regularized least squares method, LMS adaptive filter, and normalized LMS adaptive filter, and it is not necessary for the filtering result to be optimal.

[0026] The model identification unit 303 calculates autoregressive coefficients of an AR (AutoRegressive) model using the measurement data obtained when the lubricant and bearings are normal (for example, when the mileage is 10,000 km or less) as training data. The calculation of the autoregressive coefficients of the AR model is also called identification of the AR model. The model identification unit 303 calculates the autoregressive coefficients of the AR model using the training data that has passed through the optimum filter unit 302. The model identification unit 303 corresponds to an example of the model calculation unit referred to in the present invention. In this embodiment, the autoregressive coefficients of the AR model are calculated in the abnormality diagnosis device 105 equipped with the model identification unit 303. However, in the case of a diagnosis target for which the autoregressive coefficients have high versatility, the model identification unit 303 may not be provided and pre-calculated general-purpose autoregressive coefficients may be stored and used.

[0027] The model inverse filter unit 304 applies the autoregressive coefficients to the measurement data of the diagnosis target, and calculates the residual between the model and the measurement data. The model inverse filter unit 304 corresponds to an example of the residual calculation unit referred to in the present invention. The diagnosis unit 305 corresponds to an example of the determination unit referred to in the present invention, and diagnoses the deterioration of the lubricant by determining whether the state of the lubricant is normal or abnormal based on the magnitude of the residual. From the viewpoint of maintaining lubrication, it is ineffective to diagnose an abnormality only after the deterioration of the lubricant has progressed to the point where significant wear, seizure, peeling, etc. have occurred. For this reason, the diagnosis unit 305 diagnoses the deterioration of the lubricant as abnormal even when there are no scratches, etc. on the bearings or cages.

[0028] FIG. 5 is a flowchart showing the abnormality diagnosis process performed by abnormality diagnosis device 105. The abnormality diagnosis process shown in the flowchart of FIG. 5 corresponds to one embodiment of the abnormality diagnosis method of the present invention. The above-mentioned measurement data is acquired as section data consisting of time-series data for, for example, 0.5 seconds, and the abnormality diagnosis process of FIG. 5 is executed for each section data.

[0029] In step S101, measurement data is acquired as section data, and in step S102, prefilter processing is performed by the prefilter unit 301. As an example, the prefilter is implemented by an FIR (Finite Impulse Response) filter designed in the frequency domain. The prefilter needs to have a low-pass filter (LPF) function and a high-pass filter (HPF) function, but may be integrated into a band-pass filter (BPF).

[0030] If the band of the section data is limited by the pre-filter to, for example, 1 kHz or more and less than 6 kHz, and the sampling frequency of the section data is 12 kHz, the sampling frequency at the time of AD (analog-to-digital) conversion by the sensor data acquisition unit 202 is, for example, 48 kHz. This corresponds to processing in which the band is limited to less than 6 kHz by the LPF, and then 1 / 4 decimation is performed to set the sampling frequency to 12 kHz.

[0031] The prefilter processing is executed by the prefilter unit 301 as a functional unit, but may be executed by the information processing unit 201 or the sensor data acquisition unit 202 as hardware. When the prefilter processing is executed as an FIR filter, the low-pass filter processing and the decimation processing can be executed simultaneously, resulting in high calculation efficiency.

[0032] The upper limit of the frequency band in the pre-filter is desirably set to the upper limit of the flat portion of the response of the acceleration sensor 104. Also, the lower limit of the frequency band is desirably set to the limit where the effects of unnecessary mechanical vibrations are negligible.

[0033] FIG. 6 is a graph showing an example of the frequency characteristic of the response of the acceleration sensor 104. In FIG. 6, the horizontal axis indicates the frequency, and the vertical axis indicates the response of the acceleration sensor 104. The frequency characteristics indicated by the dotted line in FIG. 6 are the frequency characteristics in the case of the acceleration sensor alone, and the frequency characteristics indicated by the solid line are the frequency characteristics in the case of the acceleration sensor being attached to a machine (for example, axle box 103).

[0034] When an acceleration sensor is attached to a machine, the resonance band shifts to the lower frequency side due to contact resonance. Therefore, the upper limit of the flat part of the frequency characteristic shown by the solid line is limited to the lower frequency side compared to the frequency characteristic shown by the dotted line, as shown by arrow A in the figure. The upper limit depends on the installation method of the acceleration sensor and the condition of the installation surface.

[0035] When the upper limit of the flat portion of the frequency characteristic is 6 kHz, the upper limit of the frequency band in the pre-filter may be 6 kHz, or may be lower than 6 kHz, for example, 4 kHz.

[0036] Returning to FIG. 5, the explanation will be continued. In step S103, the section data limited to a frequency band of, for example, 1 kHz to 6 kHz by the pre-filter in step S102 is normalized with a standard deviation of 1. Then, in step S104, an optimum filter is applied by the optimum filter unit 302. In this embodiment, for example, a regularized least squares method is used as the optimum filter, but a Wiener filter, a Kalman filter, an LMS adaptive filter, or a normalized LMS adaptive filter may also be used.

[0037] The processes in steps S101 to S104 are performed on both the training data for learning the normal state and the diagnosis target data for diagnosing whether the lubrication is normal or not. If the section data is the training data (S105; yes), the model identification unit 303 identifies the AR model (calculates the autoregressive coefficients) in step S106. The autoregressive coefficients are calculated, for example, by the ordinary least squares method.

[0038] Here, the ordinary least squares method and the regularized least squares method will be described in detail. An observation vector y corresponding to interval data, which is time series data, and an unknown vector x corresponding to an autoregressive coefficient are expressed by the following equations (1) and (2).

number

number

[0039] however, m: The number of data included in the section data n: the order of the autoregressive model (m>n) k=0,1,2… The value of n is determined by trial and error using AIC (Akaike's Information Criteria) or other methods.

[0040] In the AR model (i.e., linear prediction), y and x are linearly related, so the unknown vector x and the observed vector y are linked by the relationship in the following formula (4) using the matrix H expressed by the following formula (3).

number

number

[0041] Equation (4) represents a system of simultaneous equations. Since the number of observation data m is greater than the number of unknowns n, the system of simultaneous equations does not have a solution. However, the optimal solution (least squares solution) can be found using the least squares method. On the other hand, the linear equation assumed in the case of the regularized least squares method is the following equation (5) on the assumption that noise ε is additively superimposed on the observed data y.

number

[0042] The ordinary least squares solution x(^) satisfies the normal equation (6) below.

number

[0043] Here, H T If H is regular, an appropriate inverse matrix exists, and the following equation (7) holds.

number

[0044] The least squares solution x(^) expressed by the above equation (7) can actually be found by using a numerical calculation algorithm such as QR decomposition without finding an inverse matrix. On the other hand, in the case of the observation vector y containing noise ε shown in the above formula (5), the ordinary least squares solution x(^) of formula (7) is not appropriate for filtering because it amplifies the noise. In other words, the ordinary least squares solution is overly compatible with the noise contained in the observation data, so there is no filtering effect. Therefore, H T The following equation (8) is used, in which an appropriate positive constant ξ (regularization constant) is added to the diagonal elements of H.

number

[0045] Here, I is a unit matrix. The solution x(^) expressed in the above equation (8) is the solution of the regularized least squares method, and is the optimal solution when the magnitude (absolute value) of the solution x(^) is restricted. The regularized least squares method allows the optimal solution to be found without over-fitting to noise.

[0046] By using the solution x(^) of equation (8), the observation vector (time series data) y is filtered to obtain y(^) according to the following equation (9).

number

[0047] The filter expressed by the above formula (9) acts as a Wiener filter for the interval data. The regularization constant ξ can be determined by field experience. Note that the observation vector y is assumed to have a stationary Gaussian distribution, but the actual data does not need to have an exact Gaussian distribution. For example, in the case of bearing vibration, if it is rotating without significant wear, seizure, or peeling, the interval data that expresses the vibration as discrete values ​​will be approximately Gaussian distributed.

[0048] In step S104 in FIG. 5, a time series of Gaussian distribution is separated by filtering using the regularized least squares method expressed by the above formula (9), thereby making it possible to remove noise over a wide band. In the identification of the AR model in step S106 in Fig. 5, since noise has already been removed from the training data by optimal filtering, the ordinary least squares method is used to calculate the autoregressive coefficients. That is, from y(^) obtained by filtering in the above formula (9), y and H are found again according to the above formulas (1) and (3), and the optimal solution x(^) of the autoregressive coefficients is calculated from the normal equation in the above formula (6).

[0049] The optimal solution x(^) of the autoregressive coefficient obtained by identifying the AR model differs depending on the vehicle speed and the location where the vehicle runs (i.e., the running conditions). For this reason, the speed and approximate location (bridge, elevated track, etc.) are determined as the running conditions, and in step S106, the optimal solution x(^) of the autoregressive coefficient for each running condition is calculated.

[0050] In addition, since probability fluctuations occur even under the same driving conditions, the AR model is averaged in step S107. That is, N solutions obtained from N pieces of section data under the same driving conditions are averaged using the following formula (10).

number

[0051] A specific value of N is preferably 3 or more. After averaging the AR model, the process proceeds to step S108, where a residual r0 is calculated as a residual vector in the training data obtained under normal conditions by the following formula (11).

number

[0052] Here, vector y represents training data that has been subjected to optimal filtering, and matrix H is reconstructed from the training data y after optimal filtering based on the autoregressive coefficients. The calculation to obtain residual r0 from training data y is equivalent to estimating noise ε in equation (5) above using an inverse filter, and is executed by model inverse filter unit 304. Residual r0 obtained by equation (11) is used as a criterion for determining lubrication deterioration. In the case of training data, the process of FIG. 5 ends at step S108.

[0053] On the other hand, in the case of the observation data of the diagnosis target (diagnosis target data) (S105; no), the autoregressive coefficients of the AR model calculated using the training data are used, and the residual r is calculated using the following formula (12) in step S108. i is calculated.

number

[0054] In the above formula (12), the matrix H i is the diagnostic data y i As mentioned above, the residual r i Since the calculation of is equivalent to the estimation of noise ε, the residual r calculated from the diagnostic target data in a state where lubrication deterioration does not occur is i is as small as the residual r0 in the training data. As the lubrication deterioration progresses, the residual r i Therefore, in step S109, the diagnosis unit 305 calculates the residual r i is compared to the residual r0 of the training data.

[0055] For comparing the residuals, for example, a norm (Euclidean norm) expressed by the following formula (13) is used.

number

[0056] By using the norm shown in the above formula (13), the residual vectors r0, r i It is possible to compare the magnitudes of the residual vectors. Instead of the norm, the root mean square error (RMSE) expressed by the following formula (14) may be used. Both the norm and the RMSE are faster to calculate than the absolute value average, and are convenient for comparing the magnitudes of the residual vectors.

number

[0057] Residual error r of the diagnostic data i If the residual r of the training data is larger than r0 (S109; yes), the diagnosis unit 305 determines that there is a possibility of lubrication abnormality or deterioration. i is equal to or less than the residual r0 of the training data (S109; no), the diagnosis unit 305 judges that the lubrication is within the normal range. However, since there is a certain tolerance for the deterioration of the lubricant, if the difference in the residuals is small enough not to exceed the threshold, the residuals are judged to be equal.

[0058] As described above, the pre-filter in step S102 and the optimum filter in step S104 are common processes for both the training data for learning the normal state and the training data for the diagnosis target. In particular, by applying the optimum filter in step S104 to both, the residuals r0, r i The noise ε is an appropriate estimated value, and the deterioration of the lubrication can be judged with high accuracy. In addition, even if the observation data obtained by the sensor of a relatively low frequency band is used, the deterioration of the lubrication can be judged with high accuracy.

[0059] FIG. 7 is a graph showing an example of residuals in a comparative example in which the optimum filter is omitted, and FIG. 8 is a graph showing an example of residuals in this embodiment. 7 and 8, the horizontal axis indicates the vehicle speed, and the vertical axis indicates the RMSE as the residual. In addition, in Fig. 7 and Fig. 8, the dashed line graph indicates the residual r0 of the training data, and the solid line graph indicates the residual r i This shows that.

[0060] As shown in Fig. 7, in the comparative example, the dashed line graph and the solid line graph are close to each other, and there is no clear difference in the residuals, and in some vehicle speed ranges, the residuals are reversed. In contrast, in Fig. 8, the dashed line graph and the solid line graph are clearly separated, and the residual r i is clearly larger than the residual r0 of the training data. For this reason, for example, a threshold value such as that shown in the dashed line graph can be set as the threshold value for anomaly determination.

[0061] In the above description, an example is shown in which the abnormality diagnosis device is mounted on a railway vehicle, but the abnormality diagnosis device of the present invention may be mounted on a vehicle for road use or on other types of machines having a rotating shaft, such as machine tools, other than vehicles. Alternatively, the abnormality diagnosis device of the present invention may be provided separately from a device having a rotating shaft, and the measurement signal measured by the device may be acquired by wire, wirelessly, or via a recording medium.

[0062] Furthermore, in the above description, an acceleration sensor is given as an example of a sensor that measures vibration, but the sensor that measures vibration may be a sensor other than an acceleration sensor. In the above description, the prefilter unit 301 serving both as the acquisition unit and the prefilter unit of the present invention is exemplified, but the abnormality diagnosis device of the present invention may be provided with an acquisition unit and a prefilter unit that are independent of each other. In addition, when a sensor or an amplifier located upstream of the abnormality diagnosis device outputs only a band required for abnormality diagnosis and the abnormality diagnosis device acquires the output, the abnormality diagnosis device of the present invention may be provided with only the function of the acquisition unit and not with the function of the prefilter unit. [Explanation of symbols]

[0063] 100...vehicle, 110...main body, 111...bogie, 101...wheel, 102...axle, 103...axle box, 104...acceleration sensor, 105...abnormality diagnosis device, 201...information processing unit, 202...sensor data acquisition unit, 301...pre-filter unit, 302...optimal filter unit, 303...model identification unit, 304...model inverse filter unit, 305...diagnosis unit

Claims

1. an acquisition unit that acquires vibration measurement data; a filter unit that performs filtering on the measurement data using at least one of a Wiener filter, a Kalman filter, a regularized least squares method, an LMS adaptive filter, and a normalized LMS adaptive filter; a residual calculation unit that calculates a residual by applying a coefficient of an autoregressive model to the measurement data that has passed through the filter unit; A determination unit that determines whether the residual is normal or abnormal; An abnormality diagnosis device comprising:

2. 2. The abnormality diagnosis device according to claim 1, further comprising a pre-filter section that performs filtering on the measurement data to limit a frequency band, before the filter section.

3. 2. The abnormality diagnosis device according to claim 1, further comprising a model calculation unit that calculates the autoregressive model using measurement data of vibration under normal conditions as training data.

4. The abnormality diagnosis device according to claim 3 , wherein the model calculation unit calculates the autoregressive model by using the training data that has been passed through the filter unit.

5. The abnormality diagnosis device according to claim 4 , wherein the model calculation unit calculates the autoregressive model using a least squares method.

6. An acquisition step of acquiring vibration measurement data; a filtering step of filtering the measurement data using at least one of a Wiener filter, a Kalman filter, a regularized least squares method, an LMS adaptive filter, and a normalized LMS adaptive filter; a residual calculation step of calculating a residual by applying a coefficient of an autoregressive model to the measurement data that has been subjected to the filtering step; A determination step of determining whether the residual is normal or abnormal; The abnormality diagnosis method includes the steps of:

7. A program for causing an information processing device to operate as each of the components of the abnormality diagnosis device according to any one of claims 1 to 5.

Images