Equipment diagnostic systems, equipment diagnostic methods

Abstract

To achieve diagnosis focused on equipment operating sounds and improve diagnostic accuracy.SOLUTION: A time-series signal obtaining unit 11 of an equipment diagnostic system 10 performs A / D conversion of acoustic signals of rotary machinery obtained by an acoustic sensor 6c, and periodically records the converted acoustic signals in acoustic data of an arbitrary time width. A time-frequency conversion unit 12 generates from the acoustic data an amplitude spectrogram representing time-frequency information. A feature amount calculation unit 13 calculates a feature amount using CNMF to which an initial value by NNDSVD is given based on the amplitude spectrogram. A normal information calculation unit 14 generates normal information in advance from the feature amount of a normal state of the rotary machinery, and holds the normal information. An abnormality detection unit 15 generates diagnostic information from the feature amount during diagnosis of the rotary machinery, calculates a difference between the normal information and the diagnostic information as an abnormal value, and determines that an abnormality is present when the abnormal value exceeds a threshold. A display unit 16 displays a determination result on a monitor or the like.SELECTED DRAWING: Figure 4

Description

Technical Field

[0001] The present invention relates to a technique for detecting abnormalities based on sensing data of target equipment and diagnosing the target equipment.

Background Art

[0002] In electrical equipment and production equipment, preventive maintenance is carried out to detect failures and abnormalities due to aging at an early stage. In this preventive maintenance, condition-based maintenance that performs maintenance according to the operating status and deterioration state of equipment has attracted attention.

[0003] In condition-based maintenance, while it is possible to suppress maintenance costs due to unnecessary component replacement and repair, it often relies on the experience and intuition of skilled workers in order not to miss failures and malfunctions. However, due to the shortage of skilled workers caused by the recent population decline, automation of equipment diagnosis using the following sensing data has been demanded.

[0004] First, a method of analyzing and diagnosing target equipment information from data obtained by acoustic and vibration sensors can be considered. However, actual measurement data may include not only the target equipment information but also ambient environmental noise and noise. Therefore, there is a problem that even when the equipment is operating normally, it may be erroneously judged as an abnormality due to these factors.

[0005] Next, a method of diagnosing based on the operating sound of the target equipment can also be considered. However, in an environment where there are multiple pieces of equipment, various types of operating sounds are complicatedly mixed, and there is a problem that it is impossible to determine which equipment has an abnormality.

Prior Art Documents

Patent Documents

[0006]

Patent Document 1

Non-Patent Documents

[0007] [Non-Patent Document 1] Lee, DD, Seung, HS, “Algorithms for nonnegative matrix factorization”, Advances in Neural Information Processing Systeme 13, pp. 556-562, (2000) [Non-Patent Document 2] Judith C. Brown: Calculation of a constant Q spectral transform, J. Acoust. Soc. Am., 89(1):425-434, 1991 [Non-Patent Document 3] V. Paul Pauca, J. Piper, Robert J. Plemmons, “Nonnegative matrix factorization for spectral data analysis”, Linear Algebra and its Applications 416 (2006) 29-47 [Non-Patent Document 4] C. Boutsidisa, E. Gallopoulosb, “SVD based initialization: Ahead start for nonnegative matrix factorization”, Pattern Recognition 41 (2008) 1350-1362 [Overview of the Initiative] [Problems that the invention aims to solve]

[0008] (1) As shown in Patent Document 1, a diagnostic method has been proposed that combines amplitude spectrogram and non-negative matrix factorization (NMF) based on sensing data indicating the status of the equipment. This diagnostic method targets intermittent vibration waveform data obtained from acceleration / acoustic sensors installed in equipment such as electrical facilities. After converting it into an amplitude spectrogram, it decomposes it into a basis matrix and coefficient matrix using NMF (Natural Molecular Fiber) and uses these as features to detect anomalies.

[0009] In this case, the anomaly detection method involves calculating the product of the coefficient matrix of the diagnostic data and the basis matrix previously obtained from normal data, and determining the anomaly from the difference between this product and the amplitude spectrogram in question.

[0010] (2) NMF is a dimensionality reduction technique that decomposes a non-negative matrix into two lower-rank non-negative matrices, as shown in Non-Patent Document 1. The diagnostic method in Patent Document 1, which uses NMF, may not be able to solve the aforementioned problems.

[0011] In other words, the derivation of the NMF typically involves assigning random values ​​to two matrices as initial values, and then finding the optimal solution by iteratively updating the matrix to minimize the objective function.

[0012] When NMF is applied to an amplitude spectrogram, a matrix containing frequency components and a matrix containing time-series components are obtained. As shown in Figure 1, in NMF, the "I × J (I and J are both natural numbers greater than or equal to 2)" matrix representing the amplitude spectrogram is decomposed into an "I × K matrix (hereinafter referred to as the basis matrix)" whose product is the matrix in question, and a "K × J matrix (hereinafter referred to as the coefficient matrix)".

[0013] Here, "K" is a basis number that does not exceed either "I" or "J", and is defined to approximate the original matrix. In the two matrices mentioned above, the latent elements of the original matrix appear in each basis vector (the basis matrix is ​​an I-dimensional column vector, and the coefficient matrix is ​​a J-dimensional row vector). For example, as shown in the basis matrix in Figure 2, the elements included in the amplitude spectrogram appear as peaks of different frequencies for each basis.

[0014] In the acoustic signals collected in the actual environment, in addition to the operating sounds of the target equipment, various sounds (noise) emitted from the surrounding environment are included. When applying the prior art NMF to this signal, it is assumed that multiple elements will appear in the same basis of the generated matrix due to the influence of noise and the like.

[0015] Especially when the initial value of NMF is given as random numerical values, since different values are taken for each trial, the values of the generated matrix are not uniquely determined. As a result, a diagnosis focused on the operating sound of the target equipment cannot be made, and thus there are cases where the above-mentioned problems cannot be solved.

[0016] (3) The present invention is made to solve such conventional problems, and realizes a diagnosis focused on the operating sound of the target equipment without being affected by various noises emitted from the surrounding environment of the target equipment or disturbances in the installation situation of the sensor such as when multiple equipment are operating, and aims to improve the accuracy of the diagnosis as a problem to be solved.

Means for Solving the Problems

[0017] (1) One aspect of the present invention is a system for detecting an abnormality based on vibration waveform data acquired during the operation of a target equipment and diagnosing the target equipment, a time-frequency conversion unit that generates an amplitude spectrogram representing time-frequency information from the vibration waveform data, a feature quantity calculation unit that calculates feature quantities using Constrained Non-negative Matrix Factorization (CNMF) based on the amplitude spectrogram, a normal information calculation unit that generates normal information from the feature quantities in the normal state of the target equipment in advance and holds the normal information, a abnormality detection unit that generates the feature quantities at the time of diagnosing the target equipment as diagnosis information, calculates the difference between the normal information and the diagnosis information as an abnormality value, and detects an abnormality of the target equipment according to the comparison between the abnormality value and a threshold value, and includes The feature quantity calculation unit applies CNMF as the number of bases K to the amplitude spectrogram represented by the matrix “I×J”, and calculates the basis matrix of “I×K” and the coefficient matrix of “K×J” as the feature quantities. The normal information calculation unit generates the normal information from a plurality of the basis matrices and the coefficient matrices calculated during normal operation of the target equipment. The abnormality detection unit extracts basis vectors and coefficient vectors from the same rows and columns as the normal information for the basis matrix and the coefficient matrix of the information at the time of diagnosis, and calculates the abnormal value based on these two vectors.

[0018] (2) Another aspect of the present invention is a method for detecting an abnormality and diagnosing a target equipment based on vibration waveform data acquired during operation of the target equipment using a computer, a time-frequency conversion step of generating an amplitude spectrogram representing time-frequency information from the vibration waveform data, a feature quantity calculation step of calculating feature quantities using Constrained Non-negative Matrix Factorization (CNMF) based on the amplitude spectrogram, a normal information calculation step of generating normal information from the feature quantities in the normal state of the target equipment in advance and holding the normal information, an abnormality detection step of generating the feature quantities at the time of diagnosis of the target equipment as diagnosis information, calculating the difference between the normal information and the diagnosis information as an abnormal value, and detecting an abnormality of the target equipment according to a comparison between the abnormal value and a threshold value. It has In the feature quantity calculation step, CNMF is applied as the number of bases K to the amplitude spectrum represented by the matrix “I×J”, and the basis matrix of “I×K” and the coefficient matrix of “K×J” are calculated as the feature quantities. In the normal information calculation step, the normal information is generated from a plurality of the basis matrices and the coefficient matrices calculated during normal operation of the target equipment. The anomaly detection step is characterized by extracting basis vectors and coefficient vectors from the same columns and rows as the normal information to the basis matrix and coefficient matrix of the diagnostic information, and calculating the anomaly value based on these two vectors. [Effects of the Invention]

[0019] According to the present invention, it is possible to perform a diagnosis focused solely on the operating sound of the target equipment, without being affected by various noises emitted from the surrounding environment of the target equipment or disturbances in the sensor installation situation such as the operation of multiple pieces of equipment, thereby improving the accuracy of the diagnosis. [Brief explanation of the drawing]

[0020] [Figure 1] (a) is a schematic diagram of the input matrix (amplitude spectrogram), (b) is a schematic diagram of the basis matrix decomposed by NMF, and (b) is a schematic diagram of the coefficient matrix. [Figure 2] (a) is a graph of the amplitude spectrogram, (b) is a graph showing the latent elements appearing in the basis matrix of NMF, and (c) is a graph showing the latent elements appearing in the coefficient matrix of the same matrix. [Figure 3] A diagram showing the configuration of the Target equipment in a diagnostic system according to an embodiment of the present invention. [Figure 4] Overall configuration diagram of the diagnostic system. [Figure 5] (a) is a graph showing an example of acoustic data input to the time-frequency conversion unit, and (b) is a graph showing (a) after constant Q conversion. [Figure 6] A diagram illustrating the configuration of the processing unit that operates during the learning phase. [Figure 7] A diagram illustrating the configuration of the processing unit that operates during the diagnostic phase. [Figure 8] Measurement diagram of the example [Figure 9] (a) is a graph showing the signal waveform under normal operating conditions, and (b) is a graph showing the amplitude spectrogram of (a) after constant Q transformation. [Figure 10] (a) is a graph showing the signal waveform during abnormal operation, and (b) is a graph showing the amplitude spectrogram of (a) after constant Q transformation. [Figure 11] (a) is a graph showing the basis matrix and coefficient matrix after applying CNMF to Figure 9(b), and (b) is a graph showing the basis matrix and coefficient matrix after applying CNMF to Figure 10(b). [Figure 12] An explanatory diagram showing the generation of basis vectors, coefficient vectors, and a normal spectrogram. [Figure 13] An explanatory diagram showing the process of calculating the degree of abnormality. [Figure 14] (a) is a graph showing the amplitude spectrum, (b) is a graph showing the basis matrix and coefficient matrix of NMF, and (c) is a graph showing the basis matrix and coefficient matrix of CNMF. [Figure 15] (a) is a graph showing the initial values ​​(basis matrix and coefficient matrix) of NMF, and (b) is a graph showing the updated values ​​(basis matrix and coefficient matrix) of (a). [Figure 16] (a) is a graph of the amplitude spectrogram, and (b) is a graph showing the initial values ​​(basis matrix and coefficient matrix) from NNDSVD. [Figure 17] (a) is an explanatory diagram of the relative error when no basis vectors are selected, (b) is an explanatory diagram of the relative error when basis vectors are selected, and (c) is a comparison diagram of (a) and (b). [Modes for carrying out the invention]

[0021] The following describes a diagnostic system (diagnostic method) according to an embodiment of the present invention. Figure 3 shows a rotating machinery facility as an example of equipment to be diagnosed. This rotating machinery facility 1 comprises a low-voltage motor 2, a rotating body 3, and a bearing 4, and is powered by a commercial power supply 5.

[0022] The diagnostic system detects abnormalities in the rotating machinery equipment 1 and determines whether the equipment 1 is normal or abnormal. Here, as an example, we will explain a diagnosis based on sensing data from the acoustic sensor 6c, i.e., an acoustic signal.

[0023] ≪Basic Concepts≫ (1) The diagnostic system uses the constant-Q conversion described in Non-Patent Document 2 to convert the acoustic signal into an amplitude spectrogram. This constant-Q conversion is a method of converting to frequency components by changing the number of data referenced for each frequency (time window length) so that the wavenumber is the same in the frequency band. By using the constant-Q conversion, fine frequency resolution can be obtained in the low-frequency band while simultaneously capturing rapid changes in the high-frequency band.

[0024] (2) For feature extraction of the amplitude spectrogram, we use CNMF (Constrained Non-negative Matrix Factorization) as described in Non-Patent Document 3.

[0025] CNMF is an extension of NMF that allows for more sparse representations by adding constraint terms to the objective function used to find the basis and coefficient matrices. This makes it possible to decompose the matrix into independent elements for each basis.

[0026] To explain the CNMF algorithm, as shown in equation (1), the CNMF is an amplitude spectralgram "V i、j " is a non-negative basis matrix "W i,k ", coefficient matrix "H k,j Let's consider approximating it with a linear sum of ''.

[0027]

number

[0028] Amplitude Spectrum B The squared error criterion is used to determine the degree of separation between the gram V and the matrix WH. Specifically, the basis matrix W and coefficient matrix H are found to minimize the objective function D, expressed by equation (2), while maintaining a non-negative value. These basis matrix W and coefficient matrix H are used as the features of the amplitude spectrogram. Note that "α" and "β" can be set arbitrarily.

[0029]

number

[0030] Furthermore, the basis matrix W and the coefficient matrix H can be calculated by iteratively performing the update equation shown in equation (3).

[0031]

number

[0032] (3) Non-negative double singular value decomposition (NNDSVD) is used as the initial value for CNMF.

[0033] As shown in Non-Patent Document 4, NNDSVD is a method for generating initial values ​​from an input based on singular value decomposition. If the amplitude spectrogram is a two-dimensional matrix of "I rows and J columns", then to find the initial values, singular value decomposition is performed, and then the I-dimensional column vector is obtained from the product of the first singular value and the left singular vector corresponding to that singular value. Similarly, the J-dimensional row vector is obtained from the product of the first singular value and the right singular vector corresponding to that singular value.

[0034] From the second singular value onward, the singular vectors contain negative values, so we remove them to make them non-negative while calculating the column and row vectors up to the base number K. Finally, we set an "I × K" matrix, consisting of K I-dimensional column vectors, as the initial value for the basis matrix, and a "K × J" matrix, consisting of K J-dimensional row vectors, as the initial value for the coefficient matrix.

[0035] In this case, the product of the first singular value and the singular vector corresponding to that singular value represents the component that best represents the original matrix. That is, if the component of the operating sound of the equipment in question is relatively large in the spectrogram, that component will appear in the first column (I-dimensional) of the basis matrix, and in the first row (J-dimensional) of the coefficient matrix.

[0036] If ambient noise or other noise components are larger than or equal to the operating noise, a human judgment may be used to select an I-dimensional column vector and a J-dimensional row vector containing the operating noise component from the basis matrix and coefficient matrix, respectively.

[0037] By using the I-dimensional column vectors of this basis matrix (hereinafter referred to as basis vectors) and the J-dimensional row vectors of the coefficient matrix (hereinafter referred to as coefficient vectors) as initial values, diagnosis focused on operating sounds can be achieved. Furthermore, since the initial values ​​based on singular value decomposition are uniquely determined, the aforementioned problems can be solved.

[0038] System Configuration An example of the configuration of the diagnostic system will be explained based on Figure 4. In Figure 4, 10 represents the diagnostic system.

[0039] The diagnostic system 10 is composed of a computer and is equipped with normal hardware resources (CPU, RAM, ROM, HDD, SSD, etc.). As a result of the cooperation between these hardware resources and software resources (OS, applications, etc.), the diagnostic system implements a time-series signal acquisition unit 11, a time-frequency conversion unit 12, a feature quantity calculation unit 13, a normal information calculation unit 14, an anomaly detection unit 15, and a display unit 16 as processing units.

[0040] (1) The acoustic sensor 6c uses a sound collection device such as a microphone to continuously measure the operating sound of the rotating machinery equipment 1. The operating sound measured here is input to the time-series signal acquisition unit 11 as an acoustic signal. At this time, the time-series signal acquisition unit 11 performs A / D conversion on the input acoustic signal and periodically records the converted acoustic signal as acoustic data of an arbitrary time width in a storage device such as an HDD / SSD.

[0041] (2) As shown in Figures 5(a) and 5(b), the time-frequency conversion unit 12 converts the acoustic data recorded by the time-series signal acquisition unit 11 into an amplitude spectrogram showing time-frequency information for each arbitrary time interval. Here, the constant Q conversion described in Non-Patent Document 2 is used for the conversion and generation of the amplitude spectrogram. The generated amplitude spectrogram is a two-dimensional matrix with time on the horizontal axis and frequency on the vertical axis. If the amplitude spectrogram is an "I × J" matrix, the (i,j) component represents the magnitude (sound pressure / power value) of the acoustic data.

[0042] (3) The feature calculation unit 13 calculates features based on the amplitude spectrogram generated by the time-frequency conversion unit 12. For feature extraction, the CNMF with initial values ​​generated by the aforementioned NNDSVD is used.

[0043] At this time, if CNMF is applied as the basis number K to the amplitude spectrogram represented by the matrix "I × J", the resulting matrix will consist of an "I × K" basis matrix (frequency spectrum) and a "K × J" coefficient matrix (temporal change of the frequency spectrum). The feature calculation unit 13 outputs the basis matrix and coefficient matrix as features to the normal information calculation unit 14 and the anomaly detection unit 15.

[0044] (4) The normal information calculation unit 14 learns the feature quantities extracted in advance when the rotating machinery equipment 1 is in normal operation, generates normal information (normal data / learned data) for the rotating machinery equipment 1, and stores the generated normal information in the memory device. This operation is called the learning phase (see Figure 6).

[0045] To explain the processing in the learning phase, first, the normal information calculation unit 14 receives multiple basis matrices and coefficient matrices obtained by the feature calculation unit 13 during a predetermined period of normal operation of the rotating machinery equipment 1, and normal information is calculated based on this input information.

[0046] At this time, the basis vectors in the first column are first extracted from the basis matrix, and the coefficient vectors in the first row are extracted from the coefficient matrix. This process is then performed sequentially for all basis and coefficient matrices in the normal state. Alternatively, basis vectors and coefficient vectors may be extracted from columns and rows of predetermined numbers. This extraction of both vectors is performed for all basis and coefficient matrices in the normal state.

[0047] Next, for the generated basis vectors, representative values ​​(minimum value, mean value, etc.) are calculated for each component (number) to form a single basis vector. The same process is performed on the coefficient vectors to form a single coefficient vector.

[0048] By taking the product of these two vectors, a matrix with the same magnitude as the amplitude spectrumgram is generated. This is called the normal spectrumgram, and the generated normal spectrumgram is stored in the memory as normal information. The normal information stored here is output to the abnormality detection unit 15 during diagnosis.

[0049] (5) The abnormality detection unit 15 diagnoses the rotating machinery equipment 1 using the normal information generated by the normal information calculation unit 14, i.e., the normal spectrogram. Here, the acoustic signal during the diagnosis of the rotating machinery equipment 1 is used as diagnostic data (diagnostic information), and abnormalities are detected based on the features of the diagnostic data obtained by the feature calculation unit 13. This operation is called the diagnostic phase (see Figure 7).

[0050] To explain the processing in the diagnostic phase, the anomaly detection unit 15 first extracts basis vectors and coefficient vectors from the same columns and rows as the normal information in the basis matrix and coefficient matrix of the diagnostic data. By taking the product of these two extracted vectors, a matrix of the same size as the amplitude spectrogram is generated. This is called the diagnostic spectralgram.

[0051] Next, the diagnostic spectrogram is compared with the normal spectrogram, and the difference between the two spectrograms is calculated as an anomaly value. Based on the calculated anomaly value, abnormality detection is performed on the rotating machinery equipment 1.

[0052] Here is a normal spectrogram V. n Diagnostic Spectrogram V t Therefore, the outlier E t This can be calculated using the square of the Frobenius norm in equation (4).

[0053]

number

[0054] The anomaly detection unit 15 detects the abnormal value E t If the value exceeds a pre-set threshold, the rotating machinery equipment 1 is determined to be in an abnormal state. On the other hand, abnormal value E tIf the value does not exceed the threshold, the rotating machinery equipment 1 is determined to be in a normal state. Each of these determinations is output to the display unit 16 as a diagnostic result and displayed on a monitor or the like.

[0055] Examples As an example, we will describe abnormality diagnosis using a small motor as the target equipment. In this example, foreign matter was introduced into the small motor, causing it to start making abnormal noises at a certain point in time. A microphone was also installed at the site to record the operating sound. The measurement conditions and data acquisition period at this time are as follows. • Measurement conditions Sampling size: 441,000 Sampling frequency: 44,000Hz Sampling time: 1 min Recording stop time: 9 min • Data acquisition period Normal operating conditions: 42 hours, 252 measurements. Abnormal noise level: 1 hour, 6 measurements (1) Learning Phase In the learning phase, as shown in Figure 8, a 1-minute measurement was taken every 9 minutes, and the first 6 samples (normal samples) after the start of recording were used as the acoustic signal for normal operation. First, constant Q conversion was performed on the normal samples by the time-frequency conversion unit 12, as shown in Figures 9(a) and 9(b).

[0056] As shown in Figure 11(a), the CNMF was performed on the amplitude spectrogram after the constant Q transformation by the feature calculation unit 13 to obtain the features (basis matrix and coefficient matrix). Table 1 shows the conditions for the constant Q transformation, and Table 2 shows the conditions for the CNMF.

[0057] [Table 1]

[0058] [Table 2]

[0059] Next, as shown in Figure 12, the normal information calculation unit 14 extracts the basis vectors in the first column and the coefficient vectors in the first row, which are then used as the basis vectors and coefficient vectors, respectively. Subsequently, the minimum values ​​of each component in the basis vector group and the coefficient vector group are calculated to narrow them down to one for each.

[0060] Specifically, the minimum value of each frequency bin (frequency point of the power spectrum) of the basis vector group is obtained as the normal basis vector, and the minimum value of each coefficient vector group over time is obtained as the normal coefficient vector. The normal spectrogram of normal information is generated from the product of these two vectors.

[0061] (2) Diagnostic Phase In the diagnostic phase, six samples (diagnostic samples) recorded separately from the training samples, specifically from the point of abnormal sound occurrence onward, were used. For these diagnostic samples, a constant Q transformation was first performed in the time-frequency conversion unit 12, and then the feature calculation unit 13 performed CNMF to obtain feature quantities (basis matrix and coefficient matrix).

[0062] Next, the anomaly detection unit 15 determined the basis vector and coefficient vector in the same way as the normal information calculation unit 14, then generated a diagnostic spectrogram, and calculated the degree of anomaly compared to the normal spectrogram.

[0063] The determination of the degree of abnormality will be explained based on Figure 13. In Figure 13, the horizontal axis represents the sample number, and the vertical axis represents the degree of abnormality. Sample numbers 1 to 6 represent normal samples, and sample numbers 7 to 12 represent diagnostic samples. Here, from sample number 7 onwards, the degree of abnormality increases sharply, and an abnormal increase in the margin with sample numbers 1 to 6 can be observed. If this increase exceeds the threshold, it is detected as an abnormality.

[0064] ≪Comparison Explanation≫ Based on Figures 14 to 18, the diagnostics of the diagnostic system 10 will be compared with those of the conventional technology, and its effects will be explained.

[0065] (1) Compare NMF and CNMF based on Figure 14. Here, Figure 14(a) shows the amplitude spectrograms of the two.

[0066] Figure 14(b) shows the basis matrix and coefficient matrix of NMF relative to Figure 14(a). Here, even focusing on the basis matrix, there is no difference in the frequency spectrum for each basis, and it is not possible to separate the frequency components of operating sound and disturbances (noise, etc.). Therefore, it is necessary to use the entire matrix for diagnosis, which contains many disturbances, and there is a risk that the diagnostic results will be influenced by disturbances.

[0067] On the other hand, Figure 14(c) shows the basis matrix and coefficient matrix of CNMF relative to Figure 14(a). Focusing on the basis matrix, the frequency spectra of each basis are independent, and it is possible to decompose them into independent elements for each basis.

[0068] Therefore, the frequency components of the operating noise of the rotating machinery equipment 1 and the external disturbances can be separated, and a stable diagnosis can be made by removing the external disturbance components. In this respect, the diagnostic system 10 can realize a diagnosis focused on the operating noise of the rotating machinery equipment 1, and it is expected that the accuracy of the diagnosis will be improved.

[0069] (2) Generally, the initial value of the NMF uses a random non-negative value, as shown in Figure 15(a). However, if the value of the random number changes, the initial value changes, and after updating the basis matrix and coefficient matrix to minimize the objective function (see equation (3)), the update result may change, as shown in Figure 15(b).

[0070] In contrast, the diagnostic system 10 uses NNDSVD for the initial values ​​of the basis matrix and coefficient matrix of the CNMF. For example, by applying NNDSVD to the amplitude spectrogram in Figure 16(a), the basis matrix W and coefficient matrix H in Figure 16(b) are obtained, thus resolving the problem of using random values.

[0071] Furthermore, as shown in Figure 16(b), the frequency (principal component) with the maximum intensity of the spectrogram appears in base 1 (the same applies after the update). Therefore, according to the diagnostic system 10, using NNDSVD as the initial value of CNMF allows for a more spars representation than random initial values, and an improvement in the resolution accuracy for each base can be expected.

[0072] As a result, it is possible to obtain vectors that represent independent elements for each sound source, enabling accurate identification of the operating sound as that of the rotating machinery equipment 1. This ensures that the diagnostic results are not affected by disturbances generated from the surrounding environment of the rotating machinery equipment 1, further improving the accuracy of the diagnosis.

[0073] (3) As shown in Figures 17(a) and 17(c), if no basis vector is selected, the relative error may be large and unstable. In contrast, as shown in Figures 17(b) and 17(c), selecting basis 1 results in a smaller relative error and improved diagnostic accuracy.

[0074] Other examples The present invention is not limited to the embodiments described above, and can be modified and implemented within the scope of each claim. For example, the diagnosis by the diagnostic system 10 can be applied to vibration waveform data, which is not limited to sensing data from the acoustic sensor 6c, but can be obtained by measuring the amplitude over time to obtain a time-series signal.

[0075] Therefore, sensing data obtained by measuring the input current of the low-voltage motor 2 of the rotating machinery equipment 1 using the current sensor 6a in Figure 3, or sensing data obtained by measuring the vibration of the bearing 4 of the same equipment 1 using the vibration sensor 6b, may be used. [Explanation of Symbols]

[0076] 1…Rotating machinery equipment 2... Low-voltage motor 3…Rotational body 4… Bearings 5…Commercial power supply 6a...Current sensor 6b…Vibration sensor 6c... Acoustic sensor 10…Anomaly Diagnosis System 11...Time-series signal acquisition unit 12...Time-frequency conversion section 13…Feature Calculation Unit 14…Normal information calculation unit 15... Anomaly detection unit 16…Display section

Claims

1. A system that detects abnormalities based on vibration waveform data acquired during the operation of the target equipment and diagnoses the target equipment, A time-frequency conversion unit that generates an amplitude spectrogram representing time-frequency information from the vibration waveform data, A feature calculation unit that calculates feature quantities using Constrained Non-negative Matrix Factorization (CNMF) based on the amplitude spectrogram, A normal information calculation unit generates normal information from the characteristic quantities of the normal state of the target equipment in advance and stores the normal information, An abnormality detection unit generates the characteristic quantities during the diagnosis of the target equipment as diagnostic information, calculates the difference between the normal information and the diagnostic information as an abnormal value, and detects an abnormality in the target equipment according to the comparison between the abnormal value and a threshold, Equipped with, The feature calculation unit, Applying the CNMF with a sparsity constraint to the amplitude spectrogram represented by the matrix "I × J", the basis number "K" is used to calculate the basis matrix "I × K" and the coefficient matrix "K × J" as the feature quantities. By using Nonnegative Double Singular Value Decomposition (NNDSVD) as the initial value of the CNMF, When determining the initial values, after performing singular value decomposition of the amplitude spectrogram, an I-dimensional vector is obtained from the product of the first singular value and the left singular value vector corresponding to that singular value, and a J-dimensional vector is obtained from the product of the first singular value and the right singular value vector corresponding to that singular value. From the second singular value onward, we obtain column and row vectors for a base number "K" while removing negative values ​​to ensure non-negative values. If we give an "I × K" matrix, which consists of K I-dimensional column vectors, as the initial value for the basis matrix, and a "K × J" matrix, which consists of K J-dimensional vectors, as the initial value for the coefficient matrix, The normal information calculation unit, The basis vectors, which are I-dimensional vectors in the first column, are extracted from the basis matrix of the normal state, and the coefficient vectors, which are J-dimensional vectors in the first row, are extracted from the coefficient matrix of the normal state. The normal information is then generated from the extracted basis vectors and coefficient vectors. The aforementioned abnormality detection unit, From the basis matrix and coefficient matrix at the time of the diagnosis, basis vectors and coefficient vectors in the same rows and columns as the normal state are extracted, and the diagnostic information is generated from the extracted basis vectors and coefficient vectors. A facility diagnostic system characterized by the following features.

2. The basis matrix is: The equipment diagnostic system according to claim 1, characterized in that the frequency spectrum of each base "K" is independent and can be decomposed into independent elements for each base "K".

3. The squared error criterion is adopted for the separation criterion between the amplitude spectrogram V and the matrix WH, The equipment diagnostic system according to claim 1 or 2, characterized in that it finds a basis matrix W and a coefficient matrix H that minimize the objective function D represented by equation (2) while maintaining its non-negative value. [Math 2] In equation (2), "α" and "β" can be set arbitrarily.

4. A method for detecting abnormalities and diagnosing the target equipment based on vibration waveform data acquired during the operation of the target equipment, A time-frequency conversion step of generating an amplitude spectrogram representing time-frequency information from the vibration waveform data, A feature calculation step in which features are calculated using Constrained Non-negative Matrix Factorization (CNMF) based on the amplitude spectrogram, A normal information calculation step involves generating normal information from the characteristic quantities of the normal state of the target equipment in advance and storing the normal information, An anomaly detection step includes generating the characteristic quantities during the diagnosis of the target equipment as diagnostic information, calculating the difference between the normal information and the diagnostic information as an anomaly value, and detecting an anomaly in the target equipment according to the comparison between the anomaly value and a threshold, It has, In the feature calculation step, Applying the CNMF with a sparsity constraint to the amplitude spectrogram represented by the matrix "I × J", the basis number "K" is used to calculate the basis matrix "I × K" and the coefficient matrix "K × J" as the feature quantities. By using Nonnegative Double Singular Value Decomposition (NNDSVD) as the initial value of the CNMF, When determining the initial values, after performing singular value decomposition of the amplitude spectrogram, an I-dimensional vector is obtained from the product of the first singular value and the left singular value vector corresponding to that singular value, and a J-dimensional vector is obtained from the product of the first singular value and the right singular value vector corresponding to that singular value. From the second singular value onward, we obtain column and row vectors for a base number "K" while removing negative values ​​to ensure non-negative values. If we give an "I × K" matrix, which consists of K I-dimensional column vectors, as the initial value for the basis matrix, and a "K × J" matrix, which consists of K J-dimensional vectors, as the initial value for the coefficient matrix, In the normal information calculation step, The basis vectors, which are I-dimensional vectors in the first column, are extracted from the basis matrix of the normal state, and the coefficient vectors, which are J-dimensional vectors in the first row, are extracted from the coefficient matrix of the normal state. The normal information is then generated from the extracted basis vectors and coefficient vectors. In the aforementioned anomaly detection step, From the basis matrix and coefficient matrix at the time of the diagnosis, basis vectors and coefficient vectors in the same rows and columns as the normal state are extracted, and the diagnostic information is generated from the extracted basis vectors and coefficient vectors. A method for diagnosing equipment characterized by the following features.

Images