Facility diagnostic device and facility diagnostic method

Abstract

To provide a facility diagnostic device and a method that classify, on the basis of an amount of information of various sensor data obtained from a facility to be diagnosed, an abnormality and identify a cause of the abnormality in the facility.SOLUTION: In a facility diagnostic device 8, a probability distribution conversion unit sets a distribution interval for each fixed time period for measurement data of each current, vibration, and sound sensor installed in a rotation machine facility, which is a facility to be diagnosed, divides the distribution interval into a number of classes indicating a preset random variable, and converts the distribution into a probability distribution according to the number of measurement values in each class; an information amount calculation unit calculates, on the basis of the probability distribution, the information amount between each sensor, and generates, on the basis of the calculated information amount, an information amount correlation map; a sensor network generation unit generates, on the basis of the information amount correlation map, a graph as a sensor network, in which an adjacency matrix and each sensor are a node and an edge is a component of the adjacency matrix; and a diagnostic unit diagnoses an abnormality in the rotation machine facility according to a structural change of the sensor network, and classifies a type of the abnormality according to a pattern of the sensor network.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present invention relates to a technique for diagnosing equipment to be diagnosed based on measurement data of a plurality of sensors installed in the equipment. [Background technology]

[0002] The breakdown of infrastructure facilities that support society, such as electricity and transportation, can have a large impact on society, so it is necessary to prevent unplanned shutdowns before they occur. In addition, breakdowns of production facilities in factories can also have a large impact on corporate production activities.

[0003] These facilities deteriorate with time, and there is a demand to reduce costs amid a labor shortage due to the declining birthrate and aging population, as well as increasing global competition. For this reason, unmanned and remote monitoring is expected to detect equipment failures and abnormalities early, and diagnostic techniques have been proposed that attach many sensors to the equipment to be monitored and use the multidimensional time series signals detected by each sensor.

[0004] For example, Patent Document 1 proposes that more accurate anomaly detection and failure prediction are possible by increasing the number of state quantities (physical quantities) used to determine anomalies. However, when using multiple sensor data, if the measured physical quantities are different, the characteristics of the data (units, statistics, etc.) change, making it difficult to grasp the relationships between them and making it difficult to perform correlation analysis.

[0005] Therefore, Non-Patent Document 4 proposes a method of abnormality diagnosis that is not influenced by the characteristics of data by converting the physical quantity of each sensor into information volume in correlation analysis between multiple sensor data. This method makes it possible to easily diagnose the equipment condition from the relationship between sensors by converting various different measurement data into information volume. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Patent Publication No. 2021-177074 [Non-patent literature]

[0007] [Non-Patent Document 1] S. Kullback and RA Leibler,”On information and sufficiency Ann. Math. Statist”, PP. :79-,1951 [Non-Patent Document 2] S. Amari and H. Nagaoka, “Methods of Information Geometry” A co-publication of the AMS and Oxford University Press, 2000. [Non-Patent Document 3] Takuya Ajiki, Masami Nakamura, Shunji Maeda, Takashi Komatsu, and Tetsuji Taniguchi, "Proposal of an Anomaly Detection Method Using KLD Duality," 72nd China Joint Conference of Electrical and Information Engineering Societies, 2021 [Non-Patent Document 4] Akira Imori, Shunji Maeda, Takashi Komatsu, Tetsuji Taniguchi, Osamu Tonoda, "Anomaly Detection Using the Unity, Duality, and Dependency of Entropy - Usefulness of Entropy, Kullback-Leibler Divergence, and Mutual Information -" The Institute of Electronics, Information and Communication Engineers, IEICE Technical Report Summary of the Invention [Problem to be solved by the invention]

[0008] The method of Non-Patent Document 4 is as follows:

[0009] (1) Entropy (average amount of information) of each sensor data

[0010] (2) Mutual information between sensor data

[0011] (3) "KLD (Kullback-Leibler divergence)" between the reference probability distribution and the target probability distribution These three amounts of information are used as indicators for determining abnormalities.

[0012] The entropy index (1) is applied to each sensor and its numerical change is monitored. The mutual information and "KLD" indexes (2) and (3) are applied to two different sensors or the same sensor and monitor the breakdown of the relationship between sensors and the change in distribution. Although the use of these indexes makes it possible to detect anomalies in equipment, the method in Non-Patent Document 4 does not achieve classification of anomaly types.

[0013] The present invention has been made to solve such conventional problems, and has an object to solve the problems by classifying abnormalities based on the amount of information from various sensor data obtained from the equipment to be diagnosed, and estimating the cause of the abnormality in the equipment. [Means for solving the problem]

[0014] (1) One aspect of an equipment diagnosis device according to the present invention is a device that acquires time-series measurement data of an arbitrary time width from a plurality of sensors installed in equipment to be diagnosed, and diagnoses the equipment based on the acquired measurement data, a probability distribution conversion unit that sets a distribution interval for each fixed time period for the measurement data of each of the sensors, divides the distribution interval into a number of classes that indicate a preset random variable, and converts the distribution interval into a probability distribution according to the number of measurement values ​​of each of the classes; an information amount conversion unit that calculates an information amount between each sensor based on the probability distribution and generates an information amount correlation map based on the calculated information amount; a sensor network generation unit that generates a graph as a sensor network, based on the information amount correlation map, in which an adjacency matrix and a sensor are defined as nodes and edges are defined as components of the adjacency matrix; a diagnosis unit that diagnoses the presence or absence of an abnormality in the facility in response to a structural change in the sensor network; Equipped with The diagnosis unit classifies the type of the abnormality according to a pattern of the sensor network at the time of the diagnosis.

[0015] (2) Another aspect of the equipment diagnosis device according to the present invention is a device that acquires time-series measurement data of an arbitrary time width from a plurality of sensors installed in equipment to be diagnosed, and diagnoses the equipment based on the acquired measurement data, a probability distribution conversion unit that sets a distribution interval for each fixed time period for the measurement data of each of the sensors, divides the distribution interval into a number of classes that indicate a preset random variable, and converts the distribution interval into a probability distribution according to the number of measurement values ​​of each of the classes; an information amount conversion unit that calculates an information amount between each sensor based on the probability distribution and generates an information amount correlation map based on the calculated information amount; a sensor network generation unit that generates a graph as a sensor network, based on the information amount correlation map, in which an adjacency matrix and a sensor are defined as nodes and edges are defined as components of the adjacency matrix; a matrix analysis unit that performs matrix analysis using the sensor network to obtain eigenvalues ​​that quantify the connectivity between the nodes; a diagnosis unit that diagnoses the presence or absence of an abnormality in the equipment by using the eigenvalue; An equipment diagnosis device comprising:

[0016] (3) One aspect of an equipment diagnosis method according to the present invention is a method in which a computer acquires time-series measurement data of an arbitrary time width from a plurality of sensors installed in equipment to be diagnosed, and diagnoses the equipment based on the acquired measurement data, a probability distribution conversion step of setting a distribution interval for each fixed time period for the measurement data of each sensor, dividing the distribution interval into a number of classes indicating a preset random variable, and converting the distribution interval into a probability distribution according to the number of measurement values ​​of each class; an information amount conversion step of calculating an information amount between each sensor based on the probability distribution and generating an information amount correlation map based on the calculated information amount; a sensor network generating step of generating a graph as a sensor network, the graph having an adjacency matrix and a sensor as nodes and an edge as a component of the adjacency matrix based on the information amount correlation map; a diagnosis step of diagnosing the presence or absence of an abnormality in the facility in response to a structural change in the sensor network; having The diagnosis step includes a step of classifying the type of the abnormality according to a pattern of the sensor network at the time of the diagnosis.

[0017] (4) Another aspect of the equipment diagnosis method of the present invention is a method in which a computer acquires time-series measurement data of an arbitrary time width from a plurality of sensors installed in equipment to be diagnosed, and diagnoses the equipment based on the acquired measurement data, a probability distribution conversion step of setting a distribution interval for each fixed time period for the measurement data of each sensor, dividing the distribution interval into a number of classes indicating a preset random variable, and converting the distribution interval into a probability distribution according to the number of measurement values ​​of each class; an information amount conversion step of calculating an information amount between each sensor based on the probability distribution and generating an information amount correlation map based on the calculated information amount; a sensor network generating step of generating a graph as a sensor network, the graph having an adjacency matrix and a sensor as nodes and an edge as a component of the adjacency matrix based on the information amount correlation map; a matrix analysis step of calculating eigenvalues ​​that quantify the connectivity between the nodes by performing a matrix analysis using the sensor network; and a diagnosis step of diagnosing the presence or absence of an abnormality in the equipment using the eigenvalue. Effect of the Invention

[0018] According to the present invention, since abnormalities are classified based on the amount of information of various sensor data obtained from the equipment to be diagnosed, it is possible to estimate the cause of the abnormality in the equipment. [Brief description of the drawings]

[0019] [Figure 1] 3A is a schematic diagram of the "KLD" used in the equipment diagnosis device according to the embodiment of the present invention, and FIG. 3B is a schematic diagram of the mutual information of the same. [Diagram 2] FIG. 2 is a configuration diagram showing a diagnosis situation in the first embodiment. [Diagram 3] FIG. 3 is a processing configuration diagram of the equipment diagnosis device in FIG. 2. [Figure 4] (a) is a diagram of time series data from a vibration sensor input to the probability distribution conversion unit, and (b) is a diagram of the data output after being converted to the probability distribution of (a). [Diagram 5] FIG. 1 is a configuration diagram showing an example of a sensor network. [Figure 6] FIG. 11 is a pattern diagram showing a normal state of a diagnostic example. [Figure 7] The same pattern diagram showing an abnormality caused by bearing damage. [Figure 8] (a) is a diagram of the pattern of eccentricity anomaly, and (b) is a diagram of the other pattern of the same. [Figure 9] (a) is a sensor network pattern diagram of the abnormality diagnosis result of the sensor network showing an abnormality due to bearing damage, and (b) is a sensor network pattern diagram of the abnormality diagnosis result of the sensor network showing an eccentricity abnormality. [Figure 10] A graph showing the acoustic "KLD" versus vibration. [Figure 11] A graph showing the "KLD" of current versus sound. [Figure 12] A graph showing the entropy of sound. [Figure 13] Graph showing the entropy of electric current. [Figure 14] FIG. 11 is a processing configuration diagram of an equipment diagnosis device according to a second embodiment. [Figure 15] FIG. 2 is a configuration diagram showing an example of the sensor network. [Figure 16] Graph showing the distribution of eigenvalues ​​of diagnostic example 1. [Figure 17] A graph showing the clustering results. [Figure 18] A schematic diagram of the sensor network. [Figure 19] A graph showing the classification results. [Figure 20] 3D graph showing the distribution of eigenvalue vectors of diagnostic case 2. [Figure 21] A three-dimensional graph showing the clustering results. [Figure 22] A three-dimensional graph showing the classification results. [Figure 23] 13 is a graph showing the abnormality determination results of diagnostic example 3. [Figure 24] Histogram showing the maximum eigenvalues ​​when the equipment is in normal, eccentricity abnormal, or bearing damage abnormal condition. [Diagram 25] A histogram showing the second eigenvalue. [Figure 26] A histogram showing the first component of the eigenvector of the second eigenvalue. [Figure 27] A histogram showing the second component of the eigenvector. [Figure 28] A histogram showing the third component of the eigenvector. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0020] An equipment diagnosis device (equipment diagnosis method) according to an embodiment of the present invention will be described below. This diagnosis device not only diagnoses anomalies in the equipment based on the amount of information between different sensors in the method described in Non-Patent Document 4, but also classifies the anomalies to estimate the causes of the anomalies.

[0021] The amount of information will be outlined with reference to Fig. 1. Non-Patent Document 4 converts sensor data of different physical quantities collected from a plurality of sensors attached to the equipment into a probability distribution based on occurrence frequency.

[0022] The converted probability distribution is converted to "KLD" or mutual information to unify the unit system and evaluate the correlation between sensor data. In contrast, the diagnostic device defines the relationship between sensors as a sensor network and classifies anomalies according to the patterns of each defined sensor network.

[0023] In this case, the amount of information between sensors is determined by using the "KLD" and mutual information described in Non-Patent Document 4. That is, as shown in Fig. 1(a), the "KLD" between a reference probability distribution (reference data) recorded in advance and the probability distribution of the object to be diagnosed is calculated at regular intervals.

[0024] In addition, as shown in Fig. 1(b), the mutual information I between the sensors is calculated at regular intervals. The calculated "KLD" or mutual information I is used as the amount of information between the sensors to create a sensor network, and abnormality diagnosis and classification of the equipment are performed. EXAMPLES

[0025] A first embodiment of the diagnostic device will be described with reference to Figures 2 to 13. In Figure 2, reference numeral 1 denotes a rotating machine facility (motor 1a, rotor 1b, bearing 1c) which is the facility, and reference numeral 8 denotes the diagnostic device.

[0026] Power is supplied to a motor 1a of a rotating machine equipment 1 from a commercial power source 2, and various sensors 4 to 6 are installed to monitor the equipment 1. That is, a current sensor 4 is installed on a cable that supplies power from the commercial power source to the motor 1a, and measures the input current to the motor 1a.

[0027] A vibration sensor (acceleration sensor) 5 is installed in the housing of the bearing 1c to measure the vibration of the bearing 1c. Furthermore, an acoustic sensor 6 is installed in the vicinity of the rotating machine equipment 1 to measure the operation noise of the equipment 1.

[0028] The measurement data (sensor data) obtained from the sensors 4 to 6 is input to the data measuring device 7, which aggregates and accumulates the input measurement data. That is, the data measuring device 7 includes a storage device (e.g., SSD, HDD, etc.) and stores the input measurement data of each of the sensors 4 to 6 in the storage device.

[0029] Here, the diagnostic device 8 is connected to the data measuring instrument 7, the measurement data stored in the storage device is input to the diagnostic device 8, and the result of the abnormality diagnosis of the rotating machine equipment 1 is output to a display unit 9. The display unit 9 is composed of a display device (monitor) (not shown) connected to the diagnostic device 8, and the like, and displays the diagnosis result so that the user can confirm it.

[0030] The diagnostic device 8 is configured by a computer, and includes the hardware resources of a normal computer (eg, a CPU, a RAM, a ROM, a HDD, an SSD, etc.).

[0031] As a result of the cooperation between the hardware and software resources, the diagnostic device 8 is equipped with a probability distribution calculation unit 10, an information amount conversion unit 11, a sensor network generation unit 12, and a diagnostic unit 13, as shown in FIG.

[0032] The probability distribution calculation unit 10 generates a probability distribution for each fixed time period from the measurement data of each sensor 4-6 acquired from the data measurement instrument 7, and the information amount conversion unit 11 calculates the information amount of a common unit from the probability distribution between the two heterogeneous sensors 4-6 and converts it into an information amount correlation map.

[0033] The sensor network generation unit 12 generates a sensor network indicating the relationships between the sensors 4 to 6 based on the adjacency matrix generated from the information amount correlation map, and the diagnosis unit 13 performs an abnormality diagnosis and classification of the abnormality of the rotating machine equipment 1 based on the sensor network.

[0034] <Processing steps for abnormality diagnosis> Hereinafter, a description will be given of diagnostic steps (S01 to S04: not shown) for abnormality diagnosis by the diagnostic device 8. Here, it is assumed that the sensors 4 to 6 continuously measure physical information (current value, acceleration, sound volume) of the rotating machine equipment 1, and the measured sensor data is input in advance to the data measuring device 7.

[0035] At this time, each sensor data is converted from an analog signal to a digital signal, and the converted digital signal is periodically recorded in a storage device as time-series data of an arbitrary time width. For example, the time-series data of the sensors 4 to 6 is accumulated and recorded in the data measuring device 7 every second, and the accumulated and recorded time-series data is output to the diagnostic device 8. As a result, the time-series data of the sensors 4 to 6 is input to the diagnostic device 8, and the process of the diagnostic steps S01 to S04 is started.

[0036] S 01 : When the process starts, the probability distribution calculation unit 10 generates a probability distribution for each fixed time based on each input time series data, and outputs the generated probability distribution to the information amount conversion unit 11 .

[0037] Two types of probability distributions are generated: a single sensor probability distribution and a joint probability distribution between different sensors. In this case, the time series data from each of the sensors 4 to 6 can use not only sensor data measured simultaneously, but also sensor data acquired at different times.

[0038] The probability distribution of a single sensor is calculated by dividing the possible range of the sensor data measurement values ​​(distribution interval) into small intervals (hereafter referred to as classes), determining the number of data values ​​(measurement values) included in the class interval (hereafter referred to as frequency), and dividing the frequency of each class by the total number of data (hereafter referred to as relative frequency).

[0039] In this case, it is desirable to set the distribution interval wider than the range from the maximum value to the minimum value of the measured values. Explaining this based on the sensor data of the vibration sensor 5 with reference to Figures 4(a) and (b), the time series data in Figure 4(a) is included in the measured value (±0.3), and the distribution interval is [-0.4, 0.4] (±0.4).

[0040] To find the number of measurements in each class, for example, let us define the classes as "X={x1, x2, ..., x n}" and treat this as a random variable. In this case, the probability distribution of sensor 5, "P X "(x)" is expressed as the frequency of each class divided by the number of sample data.

[0041] For example, if the order of frequencies is {0,0,0,1,6,15,33,72,...,0} and the number of sample data is "51200", then P X (x)={0,0,0,0,0.0001,0.0003,0.0006,0.0014,0.0031,0.0055,...,0}, and the probability distribution P X (x) is obtained.

[0042] By applying the same processing to the measurements of sensors 4 and 6, the probability distribution of sensors 4 and 6, P Y (x)」,「P Z This process is executed every time the time series data of the sensors 4 to 6 is input to the probability distribution calculation unit 10. As described above, since the sensor data is collected with one measurement per second, the latest probability distribution "P X (x), P Y (x), P Z (x)" is calculated.

[0043] Furthermore, the joint probability distribution indicates the joint probability between two types of sensor data from sensors 4 to 6, that is, the relative frequency when one sensor data and the other sensor data are included in each class at the same time.

[0044] S02: The information amount conversion unit 11 generates a matrix using the information amounts between two different types of sensors from the probability distribution obtained in S01, and outputs the matrix to the sensor network generation unit 12. This matrix is ​​called an information amount correlation map.

[0045] The information correlation map is generated at regular intervals by converting the probability distribution into "KLD" or mutual information. In other words, when "n" sensors are targeted, an information correlation map of size "n x n" with diagonal elements set to "0" is obtained, and the amount of information generated is " n The number of pieces will be "P2".

[0046] [Table 1] Table 1 shows the information correlation map of "KLD" when the three sensors 4 to 6 are targeted. As mentioned above, the probability distribution of sensor 5 is "P X (x)”, and the probability distribution of sensor 4 is “P Y (x)”, then “KLD” can be calculated from equation (1).

[0047]

number

[0048]

number

[0049] [Table 2] Table 2 shows the correlation map of mutual information when the three sensors 4 to 6 are targeted. The probability distribution of sensor 5 is "P X (x)”, and the probability distribution of sensor 4 is “P Y (x) and the joint probability is P X、Y If we use "(x, y)", the mutual information can be expressed in the "KLD" format and calculated using equation (3).

[0050]

number

[0051]

number

[0052] S03: The sensor network generation unit 12 generates a sensor network indicating the amount of information between sensors based on the information amount correlation map of S02, and outputs the sensor network to the diagnosis unit 13. Here, a sensor network is generated every time the information amount correlation map of S02 is generated at regular intervals.

[0053] As shown in Figure 5, a sensor network consists of a weighted directed graph in which nodes "v" and edges "e" are represented by lines with arrows. In the first place, abnormalities rarely occur in the equipment to be diagnosed, and in such cases the amount of information becomes large. A large amount of information means that the probability distributions between the two types of sensors are diverging.

[0054] The sensor network expresses the degree of such deviation with edge “e”, and when edge “e” appears, it indicates that the relationship between the sensors has been broken.

[0055] An adjacency matrix is ​​calculated when generating a sensor network. The adjacency matrix is ​​a square matrix, and its size is equal to the information amount correlation map, and each element of the adjacency matrix corresponds to the information amount correlation map. Table 3 shows the adjacency matrix when three types of sensors, such as sensors 3 to 5, are used.

[0056] [Table 3] The element "e" of the adjacency matrix represents the edge weight, and while the value of the information amount correlation map may be used as is, it may also be a value obtained by performing statistical processing such as percentage or function conversion based on each element of the information amount correlation map, or a value converted into two values ​​"0" or "1" using a threshold. If it is a binary value, the sensor network is represented by the presence or absence of an edge, and if it is a real number other than binary, the sensor network is represented by the thickness of the edge.

[0057] S04: The diagnosing unit 13 diagnoses the rotating machine equipment 1 based on the sensor network created in S03, and outputs the result of the diagnosis to the display unit 9. The contents of the diagnosis are as follows:

[0058] (1) Abnormality diagnosis to determine the presence or absence of abnormalities

[0059] (2) Classification of types of abnormalities is executed.

[0060] First, anomalies are determined by monitoring changes in the sensor network over time. If certain sensors among sensors 4 to 6 in the sensor network at a given point in time are connected to each other to represent an edge, an anomaly is determined.

[0061] Next, if an abnormality is determined, the type of abnormality is classified based on the structure of the sensor network. At this time, it is assumed that a sensor network pattern is defined in advance for each state (cause of abnormality) of the rotating machine equipment 1, and that the sensor network pattern and information indicating the state of the rotating machine equipment 1 are linked and stored in a storage device.

[0062] Therefore, at the time of diagnosis, the stored pattern is compared with the sensor network determined to be abnormal to classify the type of abnormality and estimate the cause of the abnormality in the rotating machine equipment 1.

[0063] The classified abnormality causes are output to the display unit 9 together with the results of the abnormality judgment, and are displayed on a monitor, etc. If there is no match with the pattern stored in the storage device, it is output to the display unit 9 as an abnormality that does not match.

[0064] <Diagnosis Examples> An example of a diagnosis will be described below in which the rotating machine equipment 1 in Fig. 2 is the target of diagnosis. Three types of sensors, a current sensor 4, a vibration sensor 5, and an acoustic sensor 6, are used for measurement. In the following explanation, "current", "vibration", and "acoustic" refer to the measured values ​​of the sensors 4 to 6.

[0065] In this diagnosis example, not only normal operation was diagnosed, but also two phenomena were diagnosed: abnormality due to bearing damage in the bearing 1c and abnormal eccentricity in the rotor 1b. In addition, each of the sensors 4 to 6 measured continuously for two minutes, with one sample per second. Here, measurements were taken for each phenomenon in the order of "acoustic sensor 6 → vibration sensor 5 → current sensor 4", but the diagnosis steps (S01 to S04) were executed assuming simultaneous measurements.

[0066] In this diagnosis example, the information amount correlation map "D KL " was generated.

[0067]

number

[0068] In addition, the sensor network was generated using the adjacency matrix E shown in equation (6). The elements of the adjacency matrix E are binarized values ​​as shown in equation (7), and the two variables in the same equation, "μ i,j ", "±σ i,j" indicates the threshold value for binarization. In this diagnosis example, the threshold value was calculated from an information amount correlation map of one and a half minutes (90 samples) under normal conditions.

[0069]

number

[0070]

number

[0071] On the other hand, based on the measurement data for the three minutes during which the abnormality occurred (eccentricity abnormality = one and a half minutes, bearing abnormality = one and a half minutes), the abnormality was determined to be classified into the following two patterns.

[0072] (1) During the time period when the bearing was abnormal, a sensor network (complete graph) was generated in which sensors 4 to 6 were connected in both directions, as shown in Figure 7.

[0073] (2) For the time period of eccentricity anomaly, a sensor network with one-way connections was generated, although it was not a complete graph as shown in Figures 8(a) and (b) (there is a route from acoustic sensor 6 to current sensor 4).

[0074] In this diagnosis example, since the type of abnormality was known as mentioned above, the patterns of each sensor network (patterns in Fig. 7 and Fig. 8(a) and (b)) were defined as bearing damage abnormalities and eccentricity abnormalities, and were linked to the causes of the abnormality (bearing damage and eccentricity of the rotating body 1b) and saved in a storage device.

[0075] Then, the diagnostic steps (S01 to S04) were continued for the remaining measurement data when an abnormality occurred, and as a result, only the two types of sensor networks shown in Figs. 9(a) and (b) were created.

[0076] In this case, in the sensor network of Fig. 9(a), sensors 4 to 6 are connected in both directions, similar to Fig. 7 where a bearing abnormality is defined. Also, in the sensor network of Fig. 9(b), there is a route from acoustic sensor 6 to current sensor 4, similar to Fig. 8(a) where an eccentricity abnormality is defined.

[0077] Therefore, the sensor networks in Figs. 9(a) and (b) match the patterns of the sensor networks in Figs. 7 and 8(a), and are classified as bearing anomalies and eccentricity anomalies.

[0078] In this way, the diagnostic device 8 obtains the amount of information from the measurement data of the different sensors 4 to 6. At this time, it becomes possible to detect different types of abnormalities by combining the information of the sensors 4 to 6.

[0079] For example, with the combination of "sound" and "vibration" in Fig. 10, the eccentricity abnormality of the rotor 1b has a larger margin of normality than an abnormality in a bearing, and can be positively detected. On the other hand, with the combination of "current" and "sound" shown in Fig. 11, the bearing damage abnormality has a larger margin of normality than an eccentricity abnormality, and can be positively detected.

[0080] Furthermore, by creating a sensor network from the amount of information from each of the sensors 4 to 6, it is possible to get an overview of the relationships between the sensors 4 to 6 and easily monitor the state of the rotating machine equipment 1. This provides the following effects.

[0081] (A) As shown in Fig. 12, the eccentricity abnormality of the rotating machine equipment 1 has a small entropy margin from the normal state using only "sound", and even if an abnormality can be detected, it is difficult to classify the abnormality. In contrast, with the diagnosis device 8, as shown in Fig. 10, eccentricity abnormality can be easily detected from the change in the information content (KLD) of "sound to vibration".

[0082] (B) Furthermore, abnormalities caused by bearing damage are not easy to diagnose because the change in entropy is unstable using only "current" as shown in Figure 13. In contrast, with the diagnosis device 8, abnormalities caused by bearing damage can be easily detected from the change in the information content (KLD) of "current relative to sound" as shown in Figure 11.

[0083] As a result, the diagnostic device 8 makes it possible to determine not only the presence or absence of an abnormality but also the cause of the abnormality by evaluating the correlation of the amount of information based on the measurement data of the different sensors 4 to 6 installed in the rotating machine equipment 1.

[0084] In this embodiment, the sensor network patterns defined in advance may include not only abnormalities caused by eccentricity abnormality of the rotating body 1b and bearing damage of the bearing 1c, but also the above patterns caused by other abnormalities. EXAMPLES

[0085] In the abnormality diagnosis of the embodiment, abnormalities are classified by forming a sensor network in which the relationships between the sensors 4 to 6 are graphed based on the amount of information of each sensor data obtained from the rotating machine equipment 1.

[0086] However, as the number of measurement sensors increases, the edge (amount of information) of the sensor network increases, which may make it difficult to determine anomalies. Therefore, in the second embodiment, an eigenvalue is obtained from the sensor network of the first embodiment to facilitate anomaly detection and classification.

[0087] 14, the diagnostic device 8 of the present embodiment is different from that of the first embodiment in the abnormality diagnosis by the diagnostic unit 13 since a matrix analysis unit 20 for finding the eigenvalues ​​is added. The diagnostic device 8 of the present embodiment will be described below, focusing on the processing contents of the matrix analysis unit 20 and the diagnostic unit 13.

[0088] ≪Matrix analysis section 20≫ The matrix analysis unit 20 performs matrix analysis using the sensor network (see FIG. 15) of the sensor network generation unit 12. In the matrix analysis, the Laplacian matrix, eigenvalues, eigenvectors, and spectral gap are obtained.

[0089] Here, the connectivity between nodes is quantified by finding the eigenvalues ​​of the sensor network. These eigenvalues ​​are found by solving the eigenvalue problem of the Laplacian matrix. To generate the Laplacian matrix, an order matrix and an adjacency matrix that indicates the graph structure are used. The order matrix is ​​a matrix whose diagonal components are the number of edges (order) connecting each node or the sum of the edge weights, and Table 7 shows the order matrix using three sensors 4 to 6.

[0090] [Table 7] Ingredient "d" in Table 7 xx " is "Node Vx "e" comes from the edge xy " and edge "e XZ " Similarly, the component "d yy " "d zz " can also be calculated from the edges of each node. The Laplacian matrix "L" is calculated from the degree matrix "D" in which these elements are arranged on the diagonal and the adjacency matrix "E" using equation (12). At this time, the Laplacian matrix "L" obtained by normalizing the Laplacian matrix L shown in equation (13) is obtained as W " can also be used. Here, the Laplacian matrix "L W " has all diagonal elements set to "1".

[0091]

number

[0092]

number

[0093] The spectral gap is calculated from the maximum eigenvalue and the second eigenvalue. The maximum eigenvalue indicates the largest eigenvalue among the eigenvalues ​​corresponding to the number of sensors, and the second eigenvalue indicates the second largest eigenvalue. The difference between the two eigenvalues ​​indicates the spectral gap. The connectivity of the sensor network can be grasped from this spectral gap.

[0094] The fundamental principle for generating the spectral gap is to use the difference between the maximum eigenvalue and the second eigenvalue. However, the eigenvalue for calculating the difference may be selected arbitrarily, such as using the smallest eigenvalue (minimum eigenvalue) or the second smallest eigenvalue (second eigenvalue: other than "0").

[0095] <Diagnosis Section 13> The diagnosing unit 13 performs an abnormality diagnosis on the rotating machine equipment 1 based on the Laplacian matrix, eigenvalues / eigenvectors, and spectrum gaps obtained by the matrix analyzing unit 20, and causes the display unit 9 to display the diagnosis results.

[0096] The display unit 9 displays the diagnosis results obtained from the diagnosis unit 13 on a display device such as a monitor. The diagnosis results include information on the operation status of the equipment (normal / abnormal / presence or absence of status change), the name of the abnormality, and a graph of the sensor network.

[0097] (1) Abnormality diagnosis using eigenvalues / eigenvectors The diagnosis unit 13 can perform anomaly diagnosis from the eigenvalues / eigenvectors generated by the matrix analysis unit 20. That is, machine learning or clustering techniques are applied to some of the top eigenvalues ​​or all eigenvalues ​​when sorted in descending order, or to eigenvectors of arbitrarily selected eigenvalues, or to eigenvectors of eigenvalues ​​(other than 0) sorted in descending order, based on the first component or an arbitrary component of each eigenvector.

[0098] In this case, if the possible values ​​of the eigenvalues / eigenvectors are known and correspond to the operating patterns and normal / abnormal operating states of the rotating machine equipment 1, classification is performed using supervised machine learning or clustering methods such as the k-means method or Gaussian mixture model. For machine learning, multi-class classification methods such as Support Vector Machine (SVM) and Random Forest are used.

[0099] On the other hand, when the number of clusters is unknown, clustering methods such as "Density Based Spatial Clustering of Applications with Noise (DBSCAN)" and the "X-means" method are used to automatically estimate the number of clusters.

[0100] Such a clustering method may be executed every time an eigenvalue / eigenvector is generated by the matrix analysis unit 20, or may be executed at any fixed time interval.

[0101] Each time a new cluster is formed, it is determined that a change has occurred in the operating state of the rotating machine equipment 1, and this is output to the display unit 9 together with the sensor network.

[0102] If the administrator judges that the change in status is the normal operation mode, the status information is recorded as a normal operating state. On the other hand, if the change in operating state occurs due to an abnormality, the cluster and the type of abnormality are associated and recorded, and this is the learning stage.

[0103] In the diagnosis stage, if the eigenvalues / eigenvectors found by the matrix analysis unit 20 correspond to a class in which a normality is recorded, it is determined that there is no abnormality. On the other hand, if they correspond to a cluster in which an abnormality is recorded, it is determined that the type of abnormality corresponding to that cluster has occurred, and the name of the abnormality and the sensor network at the time of the abnormality occurrence are output to the display unit 9.

[0104] (2) Abnormality diagnosis using spectrum gaps The diagnosing unit 13 can also perform an abnormality diagnosis (anomaly detection) based on the spectrum gap generated by the matrix analyzing unit 20. In this case, a threshold is set in advance from the spectrum gap when the rotating machine equipment 1 is normal, and an abnormality in the rotating machine equipment 1 is detected based on the threshold during diagnosis.

[0105] If the threshold is not exceeded, it is determined to be normal, and if the threshold is exceeded, it is determined to be abnormal, and the determination result is output to the display unit 9. The threshold may be determined by a user based on experience, or may be determined experimentally using a representative value such as the average value or standard deviation of the spectrum gap.

[0106] Instead of the spectral gap, the maximum eigenvalue, the second eigenvalue, or any other eigenvalue or eigenvector may be used for anomaly detection. Furthermore, based on the Laplacian matrix, the spectral gap, and each eigenvalue and eigenvector, it is also possible to apply change point detection methods such as the singular spectral transform method, outlier detection methods such as "One Class SVM" and "Local Outlier Factor (LOF)", and machine learning such as "Long Short-Term Memory (LSTM)" and "AutoEncoder".

[0107] In addition, whether to use "abnormality diagnosis using eigenvalues / eigenvectors" or "abnormality diagnosis using spectrum gaps" can be selected appropriately depending on the specifications and circumstances, and both may be used to diagnose abnormalities in the rotating machine equipment 1.

[0108] <Diagnosis Examples> This section explains a diagnostic example in which rotating machine equipment 1 is the equipment to be diagnosed. A current sensor 4, vibration sensor 5, and acoustic sensor 6 were used for measurement. In addition to normal operation, two phenomena are also diagnosed: abnormalities due to bearing damage and eccentricity abnormality of the rotating body 1b.

[0109] Each of sensors 4 to 6 measures each event continuously for three minutes, with one sample per second. For each event, sound, vibration, and current are measured in that order, but diagnosis processing is performed assuming simultaneous measurements. Here, 180 samples each were used for learning and diagnosis.

[0110] (1) Diagnostic example 1: "Abnormality diagnosis using eigenvalues" KLD information correlation map "D KL " was used as the adjacency matrix to generate a Laplacian matrix using equation (12). Here, the degree matrix is ​​the column-wise sum of the adjacency matrix, that is, the sum of the edge weights (KLD information amount) coming out of the node for each of sensors 4 to 6, arranged on the diagonal. The Laplacian matrix generated in this way was subjected to eigenvalue decomposition to obtain the eigenvalues ​​and eigenvectors.

[0111] In this example, we applied the clustering method "DBSCAN" to the eigenvalues. The inputs to "DBSCAN" are two variables: the largest value (maximum eigenvalue) when the eigenvalues ​​are sorted in descending order, and the next largest eigenvalue (second eigenvalue). The smallest eigenvalue (third eigenvalue) is zero, so it is not used in this example.

[0112] "DBSCAN" is a method that defines a data point as a core point when there are at least "MinPts" neighboring points, including the data point itself, within a distance "ε" of the data point, and defines points with fewer than "MinPts" neighboring points as outliers, and groups points within a distance "ε" from the core point as reachable points. The parameters of "DBSCAN" assume the L2 norm (Euclidean distance) as the inter-point distance, and set the distances to "ε=3.0, MinPts=5".

[0113] Assuming that the operating status and type of abnormality of rotating machine equipment 1 are unknown, "DBSCAN" was applied to 180 sample data for learning. Figure 16 shows the eigenvalues ​​of the extracted data, and Figure 17 shows the results of clustering. Here, the data is divided into four clusters, with all data for normality and eccentricity abnormality of rotor 1b belonging to clusters "C1" and "C2," respectively.

[0114] On the other hand, abnormalities caused by bearing scratches belong to two clusters, "C3" and "C4," but exist completely independently of other events. This makes it clear that there are four operational states, and it is possible to classify abnormalities by understanding the operational status of each cluster, "C1" to "C4."

[0115] Once the operating conditions and abnormalities in each cluster were identified, cluster "C1" was labeled as normal, cluster "C2" as an eccentricity abnormality in rotor 1b, and clusters "C3" and "C4" as abnormalities caused by bearing damage. Anomaly diagnosis was then performed on the 180 sample data for diagnosis.

[0116] Figure 18 shows an example of a sensor network for each event, with the thickness of the edges representing the amount of information between sensors 4 to 6. Figure 19 shows the diagnosis results when clustering was performed on the eigenvalues ​​of all 180 samples including these sensor networks. According to this diagnosis result, all data was properly classified, and the operating status of rotating machine equipment 1 was identified.

[0117] (2) Diagnostic example 2: "Abnormality diagnosis using eigenvectors" Based on the same data (sample) as in Diagnosis Case 1, a Laplacian matrix was generated from the "KLD" adjacency matrix, and anomaly diagnosis was performed by applying "DBSCAN" to the eigenvectors.

[0118] In this example, the eigenvector of the second eigenvalue was used. The parameters of "DBSCAN" were set as follows: "L2 norm (Euclidean distance)" was the inter-point distance, "distance ε = 0.06", and "MinPts = 5".

[0119] Figure 20 shows the eigenvectors using diagnosis example 1 and 180 learning samples. Figure 21 shows the clustering results when "DBSCAN" is applied to the eigenvectors in Figure 20, assuming that the operating status and type of abnormality of rotating machine equipment 1 are unknown. Here, the normal, eccentricity abnormality of rotor 1b, and abnormality due to bearing damage were correctly classified into clusters "C1, C2, C3" respectively.

[0120] Figure 22 shows the results of classifying 180 diagnostic samples as unknown data. As in the case of using eigenvalues, all data was properly classified, and the operating state of the rotating machine equipment 1 was identified.

[0121] (3) Diagnostic example 3: "Abnormal diagnosis due to spectrum gap" As in diagnosis example 1, "information correlation map D KL The spectral gap was generated from the eigenvalues ​​of the Laplacian matrix obtained from the following equation. In this diagnostic example, the difference between the maximum eigenvalue and the second eigenvalue was taken as the spectral gap.

[0122] Anomalies were detected using a threshold based on the average value and standard deviation of the spectrum gap under normal conditions. If the average value is "μ" and the standard deviation is "σ", then the upper limit "μ+3σ" and lower limit "μ-3σ" were used as the thresholds to detect outliers from the thresholds. 60 normal samples were used as the threshold, and data falling within the upper and lower limits was diagnosed as normal, and data outside the range was diagnosed as abnormal.

[0123] Figure 23 shows the results of anomaly diagnosis for all samples. Here, all normal data is included within the threshold range, but the eccentricity anomaly of the rotor 1b and the bearing damage anomaly are outside the threshold range. This confirmed that anomaly diagnosis can be performed by using the spectrum gap. Also, according to Figure 23, since the eccentricity anomaly has a larger margin with respect to the threshold compared to the anomaly caused by the bearing damage, it was found that each anomaly can be positively detected and the anomalies can be classified.

[0124] In this way, according to diagnosis cases 1 to 3, it was proven that by calculating the eigenvalues, eigenvectors, and spectral gaps of the sensor network consisting of the information amounts of heterogeneous sensors 4 to 6, it is possible to diagnose abnormalities in the rotating machine equipment 1 and further classify different types of abnormalities.

[0125] As a result, unlike the conventional method, all edges (n sensors = n It is not necessary to look at all P2 features, and anomaly classification is possible with fewer features (eigenvalues ​​or eigenvectors). In addition, by applying a clustering method using these eigenvalues ​​as input, anomaly classification can be achieved without setting a threshold.

[0126] Therefore, when a wide variety of measurement sensors are used, it is possible to easily perform abnormality diagnosis and abnormality classification of the rotating machine equipment 1 without adopting the method of the embodiment. Note that this embodiment is also capable of learning and diagnosing other abnormality causes in addition to abnormalities caused by eccentricity abnormality of the rotating body 1b and bearing damage of the bearing 1c.

[0127] (4) Supplementary explanation 24 to 28 will be used to provide a supplementary explanation of the abnormality classification of the rotating machine equipment 1. Fig. 24 shows a histogram of the maximum eigenvalue when the rotating machine equipment 1 is in an equipment state of normal / abnormal eccentricity of the rotor 1b / abnormality due to bearing damage, and Fig. 25 shows a histogram of the second eigenvalue when the rotating machine equipment 1 is in the same equipment state.

[0128] According to Fig. 25, the maximum eigenvalue of the abnormality due to bearing damage has a larger margin from normality than the eccentricity abnormality. On the other hand, according to Fig. 26, the second eigenvalue of the eccentricity abnormality has a larger margin from normality than the abnormality due to bearing damage. As a result, the maximum eigenvalue can positively detect the abnormality due to bearing damage, and the second eigenvalue can positively detect the eccentricity abnormality.

[0129] Moreover, Fig. 26 shows a histogram of the first component of the eigenvector of the second eigenvalue, Fig. 27 shows a histogram of the second component, and Fig. 28 shows a histogram of the third component. Fig. 26 shows that the margin between normal and two types of abnormality (abnormality due to eccentricity abnormality and bearing damage) is large, and Figs. 27 and 28 show that the margin between the two types of abnormality is large.

[0130] In this way, the margins between each phenomenon (normal, eccentricity abnormality, and abnormality due to bearing damage) also differ in the eigenvector, and by utilizing these features (margins, etc.), it becomes possible to classify abnormalities in the rotating machine equipment 1.

[0131] The present invention is not limited to the above-described embodiment, and may be modified and implemented within the scope of the claims. For example, the sensors installed in the equipment to be diagnosed are not limited to the current sensor 4, vibration sensor 5, and acoustic sensor 6, and other sensors (such as a level sensor and a flow meter) may be used. [Explanation of symbols]

[0132] 1...Rotating machine equipment (equipment to be diagnosed) 1a…Motor 1b...Rotating body 2…Commercial power supply 3...Bearing 4...Current sensor 5...Vibration sensor 6...Acoustic sensor 7. Data measuring instrument 8...Facility diagnostic equipment 9…Display section 10...Probability distribution calculation section 11...Information conversion section 12…Sensor network generation section 13. Diagnostics Department 20…Matrix analysis section

Claims

1. An apparatus for acquiring time-series measurement data of an arbitrary time width from a plurality of sensors installed in a facility to be diagnosed, and diagnosing the facility based on the acquired measurement data, comprising: a probability distribution conversion unit that sets a distribution interval for each fixed time period for the measurement data of each of the sensors, divides the distribution interval into a number of classes that indicate a preset random variable, and converts the distribution interval into a probability distribution according to the number of measurement values ​​of each of the classes; an information amount conversion unit that calculates an information amount between each sensor based on the probability distribution and generates an information amount correlation map based on the calculated information amount; a sensor network generation unit that generates a graph as a sensor network, the graph having an adjacency matrix and a sensor as nodes and an edge as a component of the adjacency matrix based on the information amount correlation map; a diagnosis unit that diagnoses the presence or absence of an abnormality in the facility in response to a structural change in the sensor network; Equipped with The diagnostic unit classifies the type of the abnormality according to a pattern of the sensor network at the time of the diagnosis. An equipment diagnosis device characterized by:

2. An apparatus for acquiring time-series measurement data of an arbitrary time width from a plurality of sensors installed in a facility to be diagnosed, and diagnosing the facility based on the acquired measurement data, comprising: a probability distribution conversion unit that sets a distribution interval for each fixed time period for the measurement data of each of the sensors, divides the distribution interval into a number of classes that indicate a preset random variable, and converts the distribution interval into a probability distribution according to the number of measurement values ​​of each of the classes; an information amount conversion unit that calculates an information amount between each sensor based on the probability distribution and generates an information amount correlation map based on the calculated information amount; a sensor network generation unit that generates a graph as a sensor network, the graph having an adjacency matrix and a sensor as nodes and an edge as a component of the adjacency matrix based on the information amount correlation map; a matrix analysis unit that performs matrix analysis using the sensor network to obtain eigenvalues ​​that quantify the connectivity between the nodes; a diagnosis unit that diagnoses the presence or absence of an abnormality in the equipment by using the eigenvalue; An equipment diagnosis device comprising:

3. The information amount conversion unit is A probability distribution of a single sensor generated by the probability distribution calculation unit; the joint probability distribution among heterogeneous sensors; The information correlation map is generated by KLD (Kullback-Leibler divergence) or mutual information using 3. The equipment diagnosis device according to claim 1 or 2.

4. The sensor network generation unit A value obtained by performing statistical processing using percentage or function conversion based on each value of the information amount correlation map generated by the information amount conversion unit, or a value obtained by converting the value into two values ​​of "0" or "1" using a threshold value, is used for the adjacency matrix.

3. The equipment diagnosis device according to claim 1 or 2.

5. The diagnosis unit includes: A pattern of the sensor network is defined in advance for each state of the equipment; Each of the defined patterns is associated with a state of the equipment and stored in a storage device, and a pattern of the sensor network at the time of diagnosis is compared with the pattern in the storage device, 2. The equipment diagnosis device according to claim 1, further comprising a step of classifying the type of the abnormality.

6. The matrix analysis unit generating a Laplacian matrix from an adjacency matrix and an order matrix indicating a structure of the sensor network, and solving the Laplacian matrix to obtain eigenvalues ​​or eigenvectors; The diagnosis unit includes: The eigenvalues ​​or the eigenvectors are clustered by learning sample data in advance, An abnormality diagnosis and an abnormality type of the equipment are classified according to a cluster to which the eigenvalue or the eigenvector at the time of diagnosis corresponds.

3. The equipment diagnosis device according to claim 2.

7. The matrix analysis unit calculates a spectral gap from a difference between the eigenvalues, while The diagnosis unit includes: If the threshold based on the normal spectrum gap is not exceeded, it is determined to be normal, and if the threshold is exceeded, it is determined to be abnormal.

3. The equipment diagnosis device according to claim 2.

8. 8. The equipment diagnosis device according to claim 7, wherein the diagnosis unit is capable of classifying types of abnormalities in the equipment according to a margin relative to the threshold value.

9. A method for diagnosing a facility to be diagnosed, comprising: a computer acquiring time-series measurement data of an arbitrary time width from a plurality of sensors installed in the facility to be diagnosed, and diagnosing the facility based on the acquired measurement data, the method comprising the steps of: a probability distribution conversion step of setting a distribution interval for each fixed time period for the measurement data of each of the sensors, dividing the distribution interval into a number of classes indicating a preset random variable, and converting the distribution interval into a probability distribution according to the number of measurement values ​​of each of the classes; an information amount conversion step of calculating an information amount between each sensor based on the probability distribution and generating an information amount correlation map based on the calculated information amount; a sensor network generating step of generating a graph as a sensor network, the graph having an adjacency matrix and a sensor as nodes and an edge as a component of the adjacency matrix based on the information amount correlation map; a diagnosis step of diagnosing the presence or absence of an abnormality in the facility in response to a structural change in the sensor network; having The diagnosis step includes a step of classifying the type of the abnormality according to a pattern of the sensor network at the time of the diagnosis. An equipment diagnosis method comprising:

10. A method for diagnosing a facility to be diagnosed, comprising: a computer acquiring time-series measurement data of an arbitrary time width from a plurality of sensors installed in the facility to be diagnosed, and diagnosing the facility based on the acquired measurement data, the method comprising the steps of: a probability distribution conversion step of setting a distribution interval for each fixed time period for the measurement data of each of the sensors, dividing the distribution interval into a number of classes indicating a preset random variable, and converting the distribution interval into a probability distribution according to the number of measurement values ​​of each of the classes; an information amount conversion step of calculating an information amount between each sensor based on the probability distribution and generating an information amount correlation map based on the calculated information amount; a sensor network generating step of generating a graph as a sensor network, the graph having an adjacency matrix and a sensor as nodes and an edge as a component of the adjacency matrix based on the information amount correlation map; a matrix analysis step of calculating eigenvalues ​​that quantify the connectivity between the nodes by performing a matrix analysis using the sensor network; a diagnosis step of diagnosing the presence or absence of an abnormality in the equipment using the eigenvalue; An equipment diagnosis method comprising:

Images