Detection device, detection method, and device for providing gravity center vector for detection

Abstract

The invention provides a detection device, a detection method, and a center-of-gravity vector providing device for detection. The operation state of an inspection object is mastered with high precision as a score through a simple structure. A center-of-gravity vector of each of N clusters obtained by clustering a plurality of state vectors (1 < = N) obtained from time-series data reflecting an operation to be inspected is stored. After the center-of-gravity vector is stored, a state vector for detection is extracted from time-series data for detection, which is newly obtained time-series data for the inspection object, and a score for the operating state of the inspection object is generated in accordance with the distance from the state vector for detection to the center-of-gravity vector.

Description

Technical Field

[0001] The present disclosure relates to a detection technique for detecting the operating state of an inspection object, a detection center-of-gravity vector providing device for providing a center-of-gravity vector required for detecting the operating state to the outside, and the like. Background Art

[0002] There are various techniques for detecting the operating state of various machines and systems as inspection objects. For example, in Patent Documents 1 and 2 below, time-series data of the normal state of an inspection object is learned in advance, a model for predicting time-series data obtained from the subsequent inspection object is generated, a failure score is generated based on the difference between the time-series data predicted by the model and the actually obtained time-series data, and an abnormality of the inspection object is detected by determining the failure score. As such a model, a machine learning model, a deep learning model, or the like is known.

[0003] Patent Document 1: Japanese Unexamined Patent Application Publication No. 2021-9441

[0004] Patent Document 2: Japanese Unexamined Patent Application Publication No. 2023-20770

[0005] However, in these methods, since the time-series data predicted by the model is compared with the time-series data from the inspection object, if the prediction accuracy of the prediction model itself is not high enough, the accuracy of the failure score is not sufficient. In order to improve the prediction accuracy, it is necessary to make the learned time-series data of a sufficient size or to make the number of layers of the neural network used for learning sufficiently large. In any case, the learning time and the hardware structure become large. Moreover, in the case where the time-series data is multi-channel (multivariate), the prediction model needs to perform predictions for all channels in principle, and the computational cost of learning and inference becomes increasingly large. Summary of the Invention

[0006] The present disclosure can be implemented in the following modes or application examples.

[0007] The present disclosure can also be implemented in the form of a detection device for detecting the operating state of an inspection object. The detection device includes: a center-of-gravity vector storage unit that stores the center-of-gravity vectors of N clusters obtained by clustering a plurality of state vectors, the plurality of state vectors being obtained from time-series data reflecting the actions of the inspection object, where 1 ≤ N; and a score generation unit that, after storing the center-of-gravity vectors, extracts a detection state vector from the time-series data at the time of detection and generates a score of the operating state of the inspection object according to the distance from the detection state vector to the center-of-gravity vector, the time-series data at the time of detection being the time-series data newly obtained for the inspection object.

[0008] In addition, the present disclosure can also be implemented as a detection method for detecting the operating state of an inspection object. In this detection method, the center-of-gravity vectors of each of the N clusters obtained by clustering and dividing a plurality of state vectors are stored, where the plurality of state vectors are obtained from time-series data reflecting the actions of the inspection object, and 1 ≤ N; after storing the center-of-gravity vectors, a detection state vector is extracted from the time-series data at the time of detection, and a score of the operating state of the inspection object is generated according to the distance from the detection state vector to the center-of-gravity vector, and the time-series data at the time of detection is the time-series data newly obtained for the inspection object.

[0009] Furthermore, the present disclosure can also be implemented as a device for providing a center-of-gravity vector for detection. The device for providing a center-of-gravity vector for detection includes: a storage unit that stores time-series data as state vectors, where the time-series data reflects the actions of an inspection object; a division unit that clusters and divides the plurality of stored state vectors into N clusters, where 1 ≤ N; a center-of-gravity vector storage unit that stores the center-of-gravity vectors obtained for each of the divided clusters; and a provision unit that provides the stored center-of-gravity vectors for comparison in detecting the operating state of the inspection object according to a request from the outside. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 is a schematic structural diagram of the anomaly detection device according to the embodiment.

[0011] Figure 2 is a schematic structural diagram of the learning mechanism.

[0012] Figure 3 is an explanatory diagram showing the initialization and learning of the reservoir and the inference points.

[0013] Figure 4 is an explanatory diagram showing the distribution of state vectors and the situation of clusters.

[0014] Figure 5 is a flowchart showing an example of the learning process.

[0015] Figure 6 is a flowchart showing an example of the processing of clusters.

[0016] Figure 7 is a flowchart showing the processing of generating a score and determining a failure score.

[0017] Figure 8 is an explanatory diagram showing the situation of determining the failure score.

[0018] Figure 9 is an explanatory diagram showing an example of the average failure score in the embodiment.

[0019] Figure 10 It is an explanatory diagram showing an example of the increase ratio of the average failure score in the embodiment.

[0020] Figure 11 It is an explanatory diagram illustrating a structure in which only the center of gravity vector storage unit is provided on the cloud side.

[0021] Figure 12 It is an explanatory diagram illustrating a structure in which the center of gravity vector storage unit and the score generation unit are provided on the cloud side.

[0022] Explanation of reference numerals

[0023] 10, 10A, 10B: Detection device; 20, 20A, 20B: Abnormality detection system; 21, 22, 23: Acceleration sensor; 30: Rotor device; 31: Base; 32: Rotating shaft; 33: Rotor; 50, 50A, 50B: Detection state vector extraction unit; 51: Input layer; 52: Reservoir layer; 60, 60A, 60B: Score generation unit; 61A, 61B: Communication unit; 70: Determination unit; 80, 80A, 80B: Center of gravity vector storage unit; 90: Output unit; 100: Learning mechanism; 110: Input layer; 120: Reservoir layer; 130: State vector storage unit; 140: Cluster division unit; 150: Center of gravity vector acquisition unit; 200: Server. Detailed description of the embodiments

[0024] A. Embodiment:

[0025] (A1) Hardware structure

[0026] The hardware structure of the detection device 10 of this embodiment will be described. Figure 1 It is a schematic structural diagram of the abnormality detection system 20 composed of the rotor device 30 and the detection device 10 for detecting its abnormality. In the figure, the learning mechanism 100 for learning for abnormality detection is shown together with the detection device 10. The learning mechanism 100 learns the data for abnormality detection before performing abnormality detection. As will be described later, if the center of gravity vector as the learning result is stored in the center of gravity vector storage unit 80, the learning mechanism 100 itself is not required for detecting the abnormality of the rotor device 30, but for the sake of convenience of explanation, it is Figure 1 shown by a dotted line in. Structural changes in the learning mechanism 100, the center of gravity vector storage unit 80, etc. will be described in detail later.

[0027] As shown in the figure, in the anomaly detection system 20, the detection device 10 detects anomalies in the rotor device 30. In this example, the detection device 10 functions as a proximity device configured to approach the rotor device 30 and detect the operating state of the rotor device 30. The rotor device 30 includes a motor MT that rotates by electricity, and is fixed to the base 31 by a first bolt A and a second bolt D. A rotor 33 is mounted on the rotating shaft 32 of the motor MT as one of the power transmission mechanisms for driving an object, such as a robotic arm. The rotor 33 is eccentrically mounted with respect to the rotating shaft 32 of the motor MT. Therefore, when the motor MT rotates, the rotor 33 disturbs the balance of the rotating shaft of the motor MT and causes it to become unbalanced. In addition, the installation states of the first bolt A and the second bolt D affect the rotation of the motor MT. The detection device 10 of the present embodiment detects as an anomaly when the first bolt A and the second bolt D are loosened, shaken, etc. and deviate from the normal range.

[0028] In the anomaly detection system 20, acceleration sensors 21, 22, 23 for the X, Y, and Z axes are provided. The acceleration sensors 21, 22, 23 for the X, Y, and Z axes detect the vibration of the rotor 33 from the three-axis directions and output signals Ux, Uy, Uz. As shown in the figure, the direction of each axis is such that the axial direction of the rotating shaft 32 of the motor MT is the Z direction, and the directions orthogonal thereto are the X direction and the Y direction. In the figure, although the acceleration sensors 21, 22, 23 are shown separately, it is also possible to use a method in which the acceleration of the three axes is detected by one sensor and acceleration signals Ux, Uy, Uz corresponding to the vibration of each axis can be output.

[0029] The detection device 10 includes: a detection state vector extraction unit 50 that inputs the signals Ux, Uy, Uz output by the acceleration sensors 21, 22, 23 and extracts a detection state vector as the state vector for anomaly detection; a score generation unit 60 that generates a score for the operating state of the rotor device 30 to be inspected; a determination unit 70 that determines the failure score of the rotor device 30 from the generated score; a centroid vector storage unit 80 that stores the centroid vector referred to when generating the score; and an output unit 90 that outputs an index corresponding to the likelihood of failure according to the magnitude of the failure score. In the detection device 10 of the present embodiment, an arithmetic logic operation circuit composed of a known CPU, GPU, memory, etc. is provided, and at least a part of the detection state vector extraction unit 50, the score generation unit 60, etc. is implemented by the CPU, etc. sequentially executing the programs stored in the memory. It should be noted that the learning mechanism 100 is the same. Of course, the detection state vector extraction unit 50, etc. can also be implemented by dedicated hardware.

[0030] The detection state vector extraction unit 50 has an input layer 51 and a reservoir layer 52 inside it. The detection state vector extraction unit 50 is configured for reservoir computing without an output layer. The input layer 51 inputs signals Ux, Uy, and Uz from the three-axis acceleration sensors 21, 22, and 23. The reservoir layer 52 has an echo state network, which is a multi-layer neural network with recurrent connections. The reservoir layer 52 outputs a detection state vector representing the state of the rotor device 30.

[0031] The extracted detection state vector is output to the score generation unit 60. The score generation unit 60 calculates a score representing the state of the rotor device 30 based on the detection state vector and the center-of-gravity vector stored in the center-of-gravity vector storage unit 80, and outputs this score. The determination unit 70 that receives the score determines the fault score of the rotor device 30. If the fault score is greater than a predetermined value, the fault score is output to the outside via the output unit 90. For the generation of the score in the score generation unit 60, a center-of-gravity vector that needs to be calculated in advance by the learning mechanism 100 and stored in the center-of-gravity vector storage unit 80 is required. Therefore, before the structure of the detection state vector extraction unit 50, the structure and function of the learning mechanism 100 will be described. It should be noted that the structure of the learning mechanism 100 and the structure of the detection state vector extraction unit 50 are the same in terms of using the function of reservoir computing. Therefore, the description of the learning mechanism 100 is also helpful for understanding the detection state vector extraction unit 50.

[0032] (A2) Structure and function of the learning mechanism 100:

[0033] Combined Figure 1 It has been described in the learning mechanism 100 shown that it is not necessarily required when detecting abnormalities during the operation of the rotor device 30. However, before the detection device 10 detects abnormalities in the rotor device 30, learning of the center-of-gravity vector and the like needs to be performed. Therefore, regarding the structure and function of the learning mechanism 100, it will be described according to Figure 2 The learning mechanism 100 is configured for reservoir computing without an output layer, and includes: an input layer 110 that inputs signals Ux, Uy, and Uz from the three-axis acceleration sensors 21, 22, and 23; a reservoir layer 120 that has an echo state network, which is a multi-layer neural network with recurrent connections; a state vector storage unit 130 that stores the output of the reservoir layer 120 as a state vector; a cluster segmentation unit 140 that divides the stored multiple state vectors into multiple clusters; and a center-of-gravity vector acquisition unit 150 that calculates the center-of-gravity vectors of the divided clusters. The center-of-gravity vectors of the respective clusters calculated by the center-of-gravity vector acquisition unit 150 are stored in the center-of-gravity vector storage unit 80.

[0034] In this embodiment, the connection weight matrix Win between the input layer 110 and the reservoir layer 120 is set to uniformly distributed random numbers in the interval [-1, 1] (random number seed: 4, scale factor: approximately 1.21). In addition, the connection weight matrix W within the reservoir layer 120 is determined as uniformly distributed random numbers in the interval [-1, 1] (random number seed: 4, synaptic density: approximately 0.4, ESP requirement: approximately 0.99 / ρ). It should be noted that ρ is the spectral radius of the matrix.

[0035] When the state vector at time t of the reservoir 120 is set to x(t) and the input data is set to u(t), the illustrated unsupervised learning model (input layer + reservoir layer) is represented by the following equation (1). x(t - 1) represents the state vector at the previous sampling time point.

[0036] x(t) = f(Win·u(t) + W·x(t - 1)…(1)

[0037] In equation (1), f refers to the activation function, and the tanh (Hyperbolic tangent function) is used here.

[0038] The input u(t) in this embodiment is the above-mentioned three-axis acceleration data Ux(t), Uy(t), Uz(t). Therefore, the dimension of the input is 3. When these acceleration data are aggregated and processed as a vector below, it is expressed as u(t). In addition, the dimension of the state vector is 64.

[0039] In this embodiment, in order to ease the update of the reservoir layer 120, the leakage rate α is used, and the state vector x(t) is obtained by the following equation (2).

[0040] x(t) = (1 - α)·x(t - 1) + α·f(Win·u(t) + W·x(t - 1))…(2)

[0041] Here, the leakage rate α is a small value around 0.0005 to 0.001.

[0042] Figure 3An example of the input data u(t) input in an unsupervised learning model composed of such an input layer 110 and a reservoir layer 120 is shown. The upper part of this figure shows an example of the case where continuous rotational vibration such as that of the rotor device 30 is used as input data, and the lower part shows an example of the case where data based on segmented operations from the start of operation to the end of product molding, such as an injection molding machine, is used. In the case of the rotor device 30, there is no clear segmentation in the three-axis acceleration signal used as input data. Therefore, the initialization INT of the reservoir layer 120 is only performed once at the beginning, and thereafter, the learning of the state vector x(t) is performed at regular intervals based on the input data u(t). It should be noted that as shown in the figure, in order to avoid the influence of the learning of the transient response of the reservoir layer 120, the timing of the first learning after initialization is delayed by a predetermined time TRNS compared to normal.

[0043] The rotational speed of the rotor device 30 to be learned is 1200 rpm, and the sampling frequency of the three-axis acceleration sensors 21, 22, and 23 that detect its vibration, i.e., acceleration, is 1 KHz. The input layer 110 downsamples the signals Ux, Uy, and Uz from the acceleration sensors 21, 22, and 23 to 0.5 KHz, which is 1 / 2 of the original, and performs the learning of the state vector at intervals of 0.5 seconds, as shown in the timing Xa1, Xa2, ….

[0044] In the case where the object of learning has a clear segmentation for each injection during operation, such as an injection molding machine, as shown in the lower part of the figure, based on the initialization INT of the state vector being performed for each cycle of the operation, the average values Xa1, Xa2, … of the state vectors during that period are obtained and learning can be performed.

[0045] Regarding the rotor device 30, when learning its vibration state, the length of each learning file is set to about 105 seconds. Considering the transient response period of the acceleration sensor and the unification of the data length, 10 seconds at the beginning of the file is removed, and the data from 10 seconds to 100 seconds, which is 90 seconds, is used.

[0046] The state vector x(t) obtained in this way is stored in the state vector storage unit 130, and the multiple stored state vectors x(t) are segmented into multiple clusters by the cluster segmentation unit 140. The cluster segmentation can be performed after storing multiple state vectors x(t), or it can be performed sequentially each time a state vector x(t) is obtained, as will be described later.

[0047] Regarding the processing of the clusters by the cluster segmentation unit 140, it will be described in detail later. Through the clusters, state vectors that are close in distance are grouped. Figure 4This situation is schematically shown in the figure. The state vector x(t) obtained through this embodiment is a multi-dimensional vector, but for the convenience of explanation in the figure, it is represented as M two-dimensional vectors. For the set BA of M state vectors x(t), the cluster segmentation unit 140 groups the state vectors x(t) that are close to each other in the region BA. In the illustrated example, the M state vectors x(t) are divided into five clusters L0 to L4. In this case, for each of the clusters L0 to L4, the centroid vector acquisition unit 150 calculates centroid vectors C0 to C5. The calculated centroid vectors are stored in the centroid vector storage unit 80.

[0048] Figure 5 An example of the processing performed by the learning mechanism 100 is shown in the flowchart. When the learning mechanism 100 starts this processing, first, an initialization process is performed (step S201). In the initialization process, the state vectors inside the input layer 110 and the reservoir layer 120 are initialized, and as described using Figure 3 the process of passing the time during which the transitional response of the rotor device 30 to be the learning object is awaited is performed. Next, the process of acquiring and saving the state vector of the learning object (step S211) is repeated a predetermined number of times (from step STR to STP). When the rotor device 30 operates and samples its three-axis accelerations Nx, Ny, and Nz at a preset sampling rate, the sampled time-series data is processed in the reservoir layer 120, and the result is obtained as the state vector x(t).

[0049] When a predetermined number of state vectors x(t) are acquired, next, cluster processing is performed (step S220). Here, SCM sequential clustering processing is performed. Figure 6 An example of this processing is shown. In Figure 5 in order to easily obtain the correspondence with the block diagram of the learning mechanism 100, it is shown that cluster processing (step S220) is performed after M state vectors x(t) are acquired, but since SCM is sequential clustering, it can be executed each time a new state vector x(t) is acquired. If after the M state vectors X(t) are stored in the state vector storage unit 130, clustering processing other than SCM, such as K-means++, hierarchical cluster segmentation method, etc., can also be adopted.

[0050] The SCM clustering process is well-known, so its detailed description is omitted. After initializing the variable i, start Figure 6The processing shown, where the variable i determines which state vector among those after the initialization process (step S201) of the state vectors to be processed. Here, the i-th state vector is denoted as x(i), and the j-th centroid vector is denoted as C(j). Here, 0 ≤ i < M and 0 ≤ j < N. When a processable state vector x(i) is input (step S221), in the case where the number of clusters is set to N, the first N (0 ≤ i < N) state vectors x(i) are directly set as the centroid vectors C(i) of the clusters (steps S222, S223). After that, the variable i is incremented (step S224), and the process is repeated from step S221 again until the processing of the M state vectors x(i) is completed (step S225: "Yes").

[0051] Starting from the (N + 1)-th state vector x(i) (step S222, "No"), the centroid vector C(p0) with the minimum distance and its distance dp0 are obtained from the centroid vectors that have already been obtained. Next, the paired centroid vectors C(p1) and C(p2) with the minimum distance and their distance dp1-p2 are obtained from the centroid vectors C (steps S222, S231, S232).

[0052] After the above processing, the magnitudes of the distance dpo and the distance dp1-p2 are discriminated. When dpo < dp1-p2, the centroid vector C(p0) and the input state vector x(i) are merged, and its value is set as the centroid vector C(p0) (steps S240, S241). On the other hand, when it is not the case that dpo < dp1-p2, the centroid vectors C(p1) and C(p2) are merged, its value is set as the centroid vector C(p1), and the state vector x(i) is set as the centroid vector C(p2) (steps S240, S242, S243). It should be noted that the merging of the vectors here is performed by weighted average with weights for the centroid vector C, and this weight takes into account the number of state vectors that have been merged. After the above processing, the increment of the variable i and whether the clustering for the M state vectors has been completed are discriminated again (steps S224, S225). If the clustering for the M state vectors is completed (step S225: "No"), the clustering process is terminated.

[0053] Through the clustering process described above, j clusters are formed from the M state vectors, as Figure 4 illustrated, and the centroid vectors C(0) to C(j - 1) of each cluster are obtained. In combination Figure 5 with this, in the successive clustering process, the calculation of the centroid vectors (step S251) and the saving of the centroid vectors (step S252) are also performed together, and the centroid vectors are saved in the centroid vector storage unit 80. If the clustering is not performed successively, as Figure 5 shown, the clustering and the calculation and saving of the centroid vectors can also be performed separately. Above. UseFigure 2 The learning institution 100 described above can also be configured separately from the detection device 10 as a gravity center vector providing device for detection and implemented independently.

[0054] (A3)Abnormality detection of the object to be inspected:

[0055] The abnormality detection process will be described, which is performed on the state of the rotor device 30 as the object to be inspected through the structure described above, that is, the learning institution 100. During the learning of the learning institution 100, M state vectors x(t) are clustered in the gravity center vector storage unit 80, and the abnormality detection process is performed in the state where the gravity center vectors C(j) of each cluster are stored. As Figure 1 shown, an input layer 51 and a reservoir layer 52 having the same structure as the input layer 110 and the reservoir layer 120 of the above-described learning institution 100 are provided in the detection state vector extraction unit 50 included in the detection device 10. Therefore, when signals of the same acceleration are input from the three-axis acceleration sensors 21, 22, and 23, the detection state vector extraction unit 50 outputs the state vector x(t) in the same manner as the learning institution 100. Based on this structure, the detection device 10 generates a score regarding the operating state of the rotor device 30 as the detection object, and if there is an abnormality in the rotor device 30, it detects this situation. Figure 7 It is a flowchart showing an example of a failure score processing routine of the abnormality detection process executed by the detection device 10.

[0056] When the detection device 10 starts this failure score processing routine in the state where learning is completed, as the rotor device 30 operates, according to the signals from the three-axis acceleration sensors 21, 22, and 23 as the detection time-series data, the state vector x(t) for detecting the operating state of the rotor device 30 is acquired (step S301). The signals from the three-axis acceleration sensors 21, 22, and 23 are the same as the signals input to the learning institution 100 during learning, but since they are not for the learning of the state vector but for detecting the operating state of the object to be inspected, for the sake of distinction, they are called "detection time-series data". Similarly, the state vector extracted by the detection state vector extraction unit 50 has the same composition as the state vector acquired by the learning institution 100 during learning, but to avoid confusion in the description, it is called "detection state vector". When the detection state vector is acquired from the operating rotor device 30, then, the process of determining the cluster L(j) existing at the position closest to this detection state vector is performed based on the clusters of the state vectors stored in the gravity center vector storage unit 80 and the gravity center vectors (step S311). By calculating the distance from each cluster to the gravity center, the closest cluster can be discriminated.

[0057] For example, as Figure 8As shown, when the acquired state vector x(t) is represented by point A, the distances (scores) from each cluster L(0) to L(4) to their respective centroid vectors C(0) to C(4) are obtained. Among such scores, the distance from the cluster L(3) with the closest distance to the centroid vector C(3) is obtained as the fault score DS (step S321).

[0058] Next, it is determined whether the fault score DS is greater than a predetermined threshold Tds (step S331). If the fault score DS is greater than the threshold Tds, processing is performed to output the fault score DS as an index indicating the degree of the fault (step S341). If the fault score DS is less than or equal to the threshold Tds, no processing is performed. As Figure 8 shown, the threshold Tds is a value that is a predetermined value greater than the average distance from each cluster L(j) to the centroid vector C(j). Therefore, in the example shown, when the average value of the predetermined number of detected state vectors x(i) is the vector reaching point A, the fault score DS corresponding to the distance from the closest cluster L(3) to the centroid vector C(3) becomes greater than the threshold Tds. In contrast, when the average value of the predetermined number of detected state vectors x(i) is the vector reaching point B, the fault score DS corresponding to the distance from the closest cluster L(0) to the centroid vector C(0) is less than or equal to the threshold Tds. Therefore, in this example, when the average value of the state vector x(i) for the predetermined number of times is the vector reaching point A, it is determined as abnormal, and some abnormal report is made from the output unit 90. This abnormal report is the output of the index of the fault in this embodiment. The output of the index of the fault can be any index or method such as voice abnormality notification, display on a printer or screen, classification of the detection object (normal product / abnormal product), etc.

[0059] Through the above processing, the abnormality detection system 20 can easily detect an abnormality generated in the rotor device 30 using the score obtained from the state vector. Moreover, when learning the signals from the three-axis acceleration sensors 21, 22, and 23 in the rotor device 30 through the learning mechanism 100 including the input layer 110 and the reservoir layer 120, there is no need to spend time repeatedly learning the weights in the multi-layer neural network. In this embodiment, since the reservoir computing method is adopted, the connection weight matrix Win between the input layer 110 and the reservoir layer 120 and the connection weight matrix W within the reservoir layer 120 only need to be set to random values in advance, and there is no need to learn the weights by methods such as backpropagation. In addition, since the state vector within the reservoir layer 120 is used, there is no need to learn the connection weight matrix leading to the output layer.

[0060] In order to verify the operation of the anomaly detection system 20 in this embodiment, the difference in the score DS under the following conditions was obtained. In the rotor device 30, the rotor 33 is eccentrically arranged with respect to the rotating shaft 32. Therefore, when the motor MT rotates, the entire rotor device 30 vibrates, and the acceleration sensors 21, 22, and 23 detect the vibration of the device. The anomaly detection system 20 needs to determine whether this vibration is normal or includes some anomaly. Figure 1 The anomalies in the rotor device 30 shown are considered to occur in states such as loosening of the first bolt and the second bolt D, and in the case of balance disorder in the normal rotor device 30. Such balance disorders are intentionally created into the following six types of imbalances.

[0061] Type (reference numeral): Content of imbalance

[0062] <1>AnDn: Rotor 33 + First bolt A + Second bolt D

[0063] <2>AnDw: Rotor 33 + First bolt A + Second bolt Dw

[0064] <3>AwDn: Rotor 33 + First bolt Aw + Second bolt D

[0065] <4>AwDw: Rotor 33 + First bolt Aw + Second bolt Dw

[0066] <5>AnDv: Rotor 33 + First bolt A + Second bolt Dv

[0067] <6>AvDn: Rotor 33 + First bolt Av + Second bolt D

[0068] It should be noted that in the above, the reference numeral w indicates that a thin washer is installed on the bolt, the reference numeral v indicates that a thick washer is installed on the bolt, and the reference numeral n indicates that no washer is installed on the bolt.

[0069] <1> to <4> respectively obtain the data of 32 files, and <5> and <6> obtain the data of 16 files. As already described, each file is a file that samples the vibration during a 90 - second period at a predetermined interval.

[0070] Figure 9 It is a chart showing an example of the average failure score in the cases of <1> to <6> above. In the figure, the inclined shading indicates the chart of the case of clustering by k - means++, and the horizontal shading indicates the chart of the case of clustering by the successive clustering method, that is, SCM. As shown in the figure, the average score of successive clustering is relatively low, but in any case of clustering, compared with the case of normally installing the first bolt A and the second bolt D (<1> AnDn: no washer), in the case of abnormal installation (<2> to <6>: with washer), the failure score is significantly higher.

[0071] Figure 10 This is clearly shown by the increase rate of the failure score. In the figure, the broken line JK represents the increase rate of the failure score in the case of clustering by k-means++, and the broken line JS represents the increase rate in the case of clustering by the successive clustering method, i.e., SCM. When taking the case of normally installing the first bolt A and the second bolt D (<1>AnDn: without washer) as a reference, it can be seen that in the case of abnormal installation (<2>~<6>: with washer), the failure score increases to about 1.2 to 1.4 times. Therefore, if the threshold Tds is appropriately selected, it is possible to easily detect the abnormality of the installation of the rotor device 30.

[0072] B. Other embodiments:

[0073] (1) The present disclosure can also be implemented in the form of a detection device for detecting the operating state of a detection and inspection object described below. The detection device includes: a center-of-gravity vector storage unit that stores the center-of-gravity vectors of N clusters obtained by clustering and dividing a plurality of state vectors, where the plurality of state vectors are obtained from time-series data reflecting the operation of the inspection object, and 1≤N; and a score generation unit that, after storing the center-of-gravity vectors, extracts a detection state vector from the detection-time time-series data and generates a score of the operating state of the inspection object according to the distance from the detection state vector to the center-of-gravity vector, where the detection-time time-series data is the newly obtained time-series data regarding the inspection object.

[0074] In this way, it is possible to extract a detection state vector from the newly obtained detection-time time-series data regarding the inspection object and generate a score of the operating state of the inspection object according to the distance from the detection state vector to the center-of-gravity vector. As the inspection object, if it is an object with clear start and end of a series of operations such as an injection molding machine or an assembly robot, the time-series data can be obtained as a unit. If it is an object without a clear starting point and ending point such as a motor or an engine, the acceleration data, etc. during a certain time range can also be obtained as a range. As the obtained time-series data, it can also be data such as acceleration, speed, position, etc., data representing vibration phenomena (amplitude, frequency, attenuation rate, resonance, etc.). If the object is based on electrical operation, it can also be electrical data representing the change along the time axis of applied voltage and current, etc. If it is an engine, a turbine, etc., time-series data such as pressure and temperature can also be used. In addition, the first-order or second-order differential values of the above data can also be used. The time-series data can be one, or it can be multi-dimensional data such as three-axis acceleration data. In addition, the time-series data can also be processed as multi-dimensional time-series data combining the above various data.

[0075] The operating state of the inspection object refers to the operating state preconceived about the state of the inspection object, such as "normal operation" or "usual operation" where there is no abnormality, or "abnormal operation" where it deviates from such operation. Of course, the operating state does not need to be divided into just two operating states of normal (usual) and abnormal. It can also be conceived as a three-way division of abnormal operating states such as normal (usual), attention required, and abnormal. In addition, the actions of the inspection object include the actions of the inspection object itself, such as the rotation of a motor, the molding action of an injection molding machine, etc. The time-series data reflecting the actions of the inspection object only needs to reflect such actions of the inspection object. The inspection object that obtains the time-series data referred to when calculating the center-of-gravity vector and the inspection object that obtains the time-series data at the time of detection when calculating the detection state vector can be the same or different inspection objects. When calculating the center-of-gravity vector from the time-series data obtained from devices of the same type, when generating a score, the time-series data at the time of detection from an inspection object different from it can also be used. For example, if a mass-produced device such as a motor is the inspection object, it is considered to use multiple samples when calculating the center-of-gravity vector, and the specific motor actually used is taken as the inspection object when calculating the score. Of course, in the case of a large and continuously used device such as an injection molding machine, time-series data can also be obtained for a specific product to calculate the center-of-gravity vector, and the detection state vector can be calculated based on the time-series data at the time of detection obtained for the same product.

[0076] The center-of-gravity vector storage unit stores the center-of-gravity vectors of each of the N clusters obtained by clustering and dividing multiple state vectors. The multiple state vectors are obtained from the time-series data reflecting the actions of the inspection object, where 1 ≤ N. The center-of-gravity vector storage unit can be prepared before the score generation of the detection device, or when the score generation of the detection device, if it is determined that the generated score is normal, the center-of-gravity vector updated using the detection state vector at the time of generating this score can be used. Since the score is generated as the distance from the cluster of the detection state vector to the center-of-gravity vector, the distance to one center-of-gravity vector can be processed as a scalar, or the distances to multiple center-of-gravity vectors can be processed as a vector of two dimensions or more.

[0077] (2) In the above structure, the score generation unit may include a determination unit that determines the minimum distance among the distances from the detection state vector to the N centroid vectors as the failure score of the inspection object. In this way, when the inspection object is abnormal, it is easy to obtain a failure score that reflects this state. By generating or determining the above scores and failure scores, the state of the inspection object can be traced. For example, when the inspection object appears on the market and a failure or other abnormality occurs during or after the operation guarantee period, if the scores and failure scores are obtained, tracking these scores can verify the relationship between the initial condition and the subsequent changes of the inspection object. Even for an inspection object without a failure or other abnormality, the relationship between the scores and failure scores and the subsequent soundness of the inspection object can be verified.

[0078] (3) In the structures of the above (1) and (2), an output unit may be provided. When the failure score is equal to or greater than a pre-determined value, the output unit outputs an index corresponding to the probability of failure of the inspection object. In this way, it is possible to know the probability that the inspection object is in a state leading to failure and take corresponding measures. For example, it is possible to perform pre-shipment correction, repair, reassembly, classification, etc.

[0079] (4) In the structures of the above (1) to (3), the detection device may include the centroid vector storage unit and the score generation unit in a proximity device. The proximity device is arranged to be close to the inspection object and directly input the time-series data reflecting the operation of the inspection object. In this way, it is possible to immediately refer to the centroid vector and easily generate scores. In the above first embodiment, the detection device 10 is configured as a proximity device that is arranged close to the rotor device 30 and directly inputs time-series data. Additionally, depending on the situation, clustering and centroid vector calculations may be performed based on the state vector obtained from the input time-series data during detection, and the storage content of the centroid vector storage unit may be updated with the calculated new centroid vector.

[0080] (5) In the structures of the above (1) to (3), the detection device may include the score generation unit in a proximity device. The proximity device is arranged to be close to the inspection object and directly input the time-series data during detection. The detection device includes the centroid vector storage unit in the cloud, where the cloud is connected to the proximity device via a network and can exchange data with the proximity device. In this way, high-load processing such as the generation of centroid vectors can be placed on the cloud side, and the structure inside the proximity device can be simplified.

[0081] Figure 11The structure of such an anomaly detection system 20A is illustrated. A detection device 10A as a proximity device is connected via a network NW such as the Internet to a server 200 placed in the cloud, and can access a centroid vector storage unit 80A in a storage medium 210 such as a hard disk in the server 200. The detection device 10A inputs signals from triaxial acceleration sensors 21, 22, and 23 of a rotor device 30, and outputs a detection state vector through a detection state vector extraction unit 50A. A score generation unit 60A that receives the detection state vector generates a score from the communication unit 61A via the network NW with reference to the centroid vector storage unit 80A of the server 200 according to the distance to the centroid vector stored therein. Other structures are the same as those of the first embodiment.

[0082] In the structures of (1) to (3) above, the detection device may include a proximity device that is arranged to be close to the inspection object and directly inputs the detection time series data. The detection device may include the centroid vector storage unit and the score generation unit in the cloud. The cloud is connected to the proximity device via a network and can perform data exchange with the proximity device. In this way, high-load processes such as the generation of centroid vectors and the generation of scores can be placed on the cloud side, and the structure within the proximity device can be simplified.

[0083] Figure 12 The structure of such an anomaly detection system 20B is illustrated. A detection device 10B as a proximity device is connected via a network NW such as the Internet to a server 200 placed in the cloud, and can output the detection state vector extracted by a detection state vector extraction unit 50B to a score generation unit 60B provided in the server 200 via a communication unit 61B and the network NW. The score generation unit 60B in the server 200B generates a score with reference to the centroid vector storage unit 80B in a storage medium 210 such as a hard disk according to the distance to the centroid vector stored therein. The generated score is returned to the detection device 10B via the network NW and the communication unit 61B. Other structures are the same as those of the first embodiment.

[0084] The present disclosure can also be implemented as a detection method for detecting the operating state of an inspection object. In this detection method, the centroid vectors of each of the N clusters obtained by clustering and dividing a plurality of state vectors are stored. The plurality of state vectors are obtained from time-series data reflecting the actions of the inspection object, where 1 ≤ N. After storing the centroid vectors, a state vector for detection is extracted from the time-series data at the time of detection, and a score of the operating state of the inspection object is generated according to the distance from the state vector for detection to the centroid vector. The time-series data at the time of detection is the newly obtained time-series data regarding the inspection object. According to this detection method, a state vector for detection can be extracted from the newly obtained time-series data at the time of detection regarding the inspection object, and a score of the operating state of the inspection object can be generated according to the distance from the state vector for detection to the centroid vector. Regarding the inspection object, etc., it is the same as the inspection device in (1) above.

[0085] (8) The present disclosure can also be implemented as a device for providing centroid vectors for detection. The device for providing centroid vectors for detection includes: a storage unit that stores time-series data as state vectors, the time-series data reflecting the actions of an inspection object; a division unit that clusters and divides the plurality of stored state vectors into N clusters, where 1 ≤ N; a centroid vector storage unit that stores the centroid vectors obtained for each of the divided clusters; and a provision unit that, according to a request from the outside, provides the stored centroid vectors for use in comparing the operating state of the inspection object. In this way, it is possible to easily provide centroid vectors to an external device that detects the operating state of an inspection object. The centroid vectors are obtained by dividing a plurality of state vectors reflecting the actions of the inspection object into N clusters and calculating the centroid vectors for each cluster.

[0086] (9) In the above structure, it may be that the storage unit includes a multi-layer recurrent neural network that inputs the time-series data and stores the output of the neural network as the state vector. In this way, it is possible to easily process the time-series data and simplify the storage of the state vector.

[0087] (10) In the above structure, it may be that the storage unit operates the neural network by an algorithm of ESN (Echo State Network) or LSM (Liquid State Machine).

[0088] (11) In each of the above embodiments, a part of the structure implemented by hardware may be replaced with software. At least a part of the structure implemented by software can also be implemented by a discrete circuit structure. In addition, when a part or all of the functions of the present disclosure are implemented by software, the software (computer program) can be provided in a form stored in a computer-readable recording medium. The "computer-readable recording medium" is not limited to portable recording media such as floppy disks and CD-ROMs, but also includes various internal storage devices in a computer such as various RAMs and ROMs, and external storage devices fixed to the computer such as hard disks. That is, the "computer-readable recording medium" has a broad meaning including any recording medium capable of non-temporarily fixing data packets.

[0089] The present disclosure is not limited to the above embodiments and can be implemented by various structures within the scope of not departing from its gist. For example, in order to solve a part or all of the above technical problems, or achieve a part or all of the above effects, the technical features in the embodiments corresponding to the technical features in each of the embodiments described in the summary of the invention can be appropriately replaced and combined. In addition, if the technical feature is not an essential feature to be described in this specification, it can be appropriately deleted.

Claims

1. A detection device, characterized in that, For detecting the operating state of an inspection object, the detection device includes: A center-of-gravity vector storage unit that stores the center-of-gravity vectors of N clusters obtained by clustering multiple state vectors, where the multiple state vectors are obtained from time-series data reflecting the actions of the inspection object, and 1 ≤ N; and A score generation unit that, after storing the center-of-gravity vectors, extracts a detection state vector from the time-series data at the time of detection, and generates a score of the operating state of the inspection object according to the distance from the detection state vector to the center-of-gravity vector, where the time-series data at the time of detection is the newly obtained time-series data regarding the inspection object.

2. The detection device according to claim 1, wherein: The score generation unit includes a determination unit that determines the minimum distance among the distances from the detection state vector to the N center-of-gravity vectors as the failure score of the inspection object.

3. The detection device according to claim 2, wherein: The detection device includes an output unit that outputs an index corresponding to the possibility of failure of the inspection object when the failure score is equal to or greater than a predetermined magnitude.

4. The detection device according to any one of claims 1 to 3, wherein: The detection device has the center-of-gravity vector storage unit and the score generation unit in a proximity device, and the proximity device is arranged to be close to the inspection object and directly input the time-series data reflecting the actions of the inspection object.

5. The detection device according to any one of claims 1 to 3, wherein: The detection device has the score generation unit in a proximity device, and the proximity device is arranged to be close to the inspection object and directly input the time-series data at the time of detection, The detection device has the center-of-gravity vector storage unit in the cloud, and the cloud is connected to the proximity device via a network and can perform data exchange with the proximity device.

6. The detection device according to any one of claims 1 to 3, wherein: The detection device includes a proximity device that is arranged to be close to the inspection object and directly input the time-series data at the time of detection, The detection device has the center-of-gravity vector storage unit and the score generation unit in the cloud, and the cloud is connected to the proximity device via a network and can perform data exchange with the proximity device.

7. A detection method, characterized in that, For detecting the operating state of an inspection object, Storing the center-of-gravity vectors of N clusters obtained by clustering multiple state vectors, where the multiple state vectors are obtained from time-series data reflecting the actions of the inspection object, and 1 ≤ N, After storing the center-of-gravity vectors, extracting a detection state vector from the time-series data at the time of detection, and generating a score of the operating state of the inspection object according to the distance from the detection state vector to the center-of-gravity vector, where the time-series data at the time of detection is the newly obtained time-series data regarding the inspection object.

8. A center of gravity vector providing device for detection, characterized in that, Includes: A storage unit that stores time-series data as state vectors, where the time-series data reflects the actions of an inspection object; A dividing unit that divides the accumulated multiple state vector clusters into N clusters, where 1 ≤ N; A centroid vector storage unit that stores the centroid vectors obtained for each of the divided clusters; and A providing unit that provides the stored centroid vectors for comparison in detecting the operating state of the inspection object according to a request from the outside.

9. The centroid vector providing device for detection according to claim 8, characterized in that The accumulating unit includes a multi-layer recursive neural network that inputs the time-series data, and accumulates the output of the neural network as the state vector.

10. The centroid vector providing device for detection according to claim 9, characterized in that The accumulating unit operates the neural network by an algorithm of an echo state network or a liquid state machine.

Images