Detector, detection method, and providing device of center-of-gravity vector for detection

Abstract

To accurately grasp an operational state of a test object as a score with a simple configuration.SOLUTION: A center-of-gravity vector of each of N (1≤N) clusters is memorized, the clusters being divided from multiple state vectors obtained from time-series data reflecting an operation of a test object. A state vector for detection is extracted from the time-series data at the time of detection that are the time-series data newly obtained with respect to the test object after the center-of-gravity vector is memorized, and a score of an operational state of the test object is generated according to a distance from the center-of-gravity vector of the state vector for detection.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a detection technology for detecting the driving state of an object to be inspected, and a detection centroid vector providing device for externally providing a centroid vector required for detecting the driving state. [Background technology]

[0002] Various technologies have been proposed for detecting the operating status of various machines and systems as inspection targets. For example, in Patent Documents 1 and 2 listed below, time series data of the inspection target in its normal state is learned in advance, a model is generated that predicts time series data acquired from the inspection target thereafter, a fault score is generated from the difference between the time series data predicted by the model and the actually acquired time series data, and an abnormality in the inspection target is detected by judging the fault score. Known examples of such models include machine learning models and deep learning models. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2021-9441 [Patent Document 2] Japanese Patent Publication No. 2023-20770 Summary of the Invention [Problem to be solved by the invention]

[0004] However, these methods compare the time series data predicted by the model with the time series data from the test object, so unless the predictive accuracy of the predictive model itself is high enough, the accuracy of the fault score will not be sufficient. To improve the predictive accuracy, it is necessary to make the time series data to be learned sufficiently large or to make the number of layers of the neuron network for learning sufficiently large. In either case, the time required for learning and the hardware configuration become long. Moreover, when the time series data is multi-channel (multivariate), the predictive model must, in principle, predict all channels, which increases the computational cost of learning and inference. [Means for solving the problem]

[0005] The present disclosure can be realized in the following forms or application examples.

[0006] The present disclosure can also be implemented as a detection device for detecting a driving state of an object to be inspected. The detection device includes a centroid vector storage unit that stores centroid vectors for each of N (1≦N) clusters obtained by cluster-dividing a plurality of state vectors obtained from time-series data reflecting the behavior of the object to be inspected, and a score generation unit that, after the centroid vectors are stored, extracts a detection state vector from detection-time time-series data that is newly obtained time-series data regarding the object to be inspected, and generates a score for the driving state of the object to be inspected according to a distance of the detection state vector from the centroid vector.

[0007] The present disclosure can also be implemented as a detection method for detecting a driving state of an object to be inspected. In this detection method, a plurality of state vectors obtained from time-series data reflecting the behavior of the object to be inspected are cluster-divided into N (1≦N) clusters, and a centroid vector for each of the clusters is stored. After the centroid vectors are stored, a detection state vector is extracted from detection-time time-series data, which is the time-series data newly obtained for the object to be inspected, and a score of the driving state of the object to be inspected is generated according to the distance of the detection state vector from the centroid vector.

[0008] Furthermore, the present disclosure can be implemented as a device for providing a centroid vector for detection, which includes an accumulation unit that accumulates time-series data reflecting the operation of an object to be inspected as a state vector, a division unit that divides the accumulated state vectors into N (1≦N) clusters, a centroid vector storage unit that stores a centroid vector calculated for each of the divided clusters, and a provision unit that, in response to an external request, provides the stored centroid vectors for comparison to detect the operating state of the object to be inspected. [Brief explanation of the drawings]

[0009] [Figure 1] 1 is a schematic configuration diagram of an abnormality detection device according to an embodiment; [Figure 2] Schematic diagram of the learning mechanism. [Figure 3] Schematic showing reservoir initialization and training and inference points. [Figure 4] FIG. 10 is an explanatory diagram showing the distribution of state vectors and clustering. [Figure 5] 10 is a flowchart showing an example of a learning process. [Figure 6] 10 is a flowchart showing an example of a clustering process. [Figure 7] 10 is a flowchart showing a process for generating scores and identifying fault scores. [Figure 8] FIG. 10 is an explanatory diagram showing how a failure score is specified. [Figure 9] FIG. 10 is an explanatory diagram illustrating an example of an average failure score according to the embodiment. [Figure 10] FIG. 10 is an explanatory diagram showing an example of an increase ratio of the average failure score in the embodiment. [Figure 11] FIG. 10 is an explanatory diagram illustrating a configuration in which only a centroid vector storage unit is provided on the cloud side. [Figure 12] FIG. 10 is an explanatory diagram illustrating a configuration in which a centroid vector storage unit and a score generation unit are provided on the cloud side. DETAILED DESCRIPTION OF THE INVENTION

[0010] A. Implementation: (A1) Hardware configuration: The hardware configuration of the detection device 10 of this embodiment will be described. FIG. 1 is a schematic diagram of an anomaly detection system 20 consisting of a rotor device 30 and a detection device 10 that detects anomalies in the rotor device 30. The diagram also shows a learning mechanism 100 that performs learning for anomaly detection, along with the detection device 10. The learning mechanism 100 learns data for anomaly detection before performing anomaly detection. As will be described later, once the centroid vectors resulting from the learning are stored in the centroid vector storage unit 80, the learning mechanism 100 itself is not necessary for detecting anomalies in the rotor device 30. However, for convenience of explanation, the learning mechanism 100 is shown in FIG. 1 with a dashed line. Variations in the configuration of the learning mechanism 100, centroid vector storage unit 80, and the like will be described in detail later.

[0011] As shown in the figure, in the anomaly detection system 20, the target for which the detection device 10 detects an anomaly is a rotor device 30. In this example, the detection device 10 is disposed adjacent to the rotor device 30 and functions as a proximity device that detects the operating state of the rotor device 30. The rotor device 30 includes a motor MT that rotates using electricity and is fixed to a base 31 with a first bolt A and a second bolt D. A rotor 33 is attached to a rotation shaft 32 of the motor MT as part of a power transmission mechanism for driving an object, such as a robot arm. The rotor 33 is attached eccentrically to the rotation shaft 32 of the motor MT. Therefore, when the motor MT rotates, the rotor 33 disrupts the balance of the rotation shaft of the motor MT, causing it to become unbalanced. Furthermore, the attachment state of the first bolt A and the second bolt D affects the rotation of the motor MT. The detection device 10 of this embodiment detects as an abnormality when the looseness or rattle of the first bolt A or the second bolt D is outside the normal range.

[0012] The abnormality detection system 20 is provided with X-, Y-, and Z-axis acceleration sensors 21-23 that detect vibrations of the rotor 33 in each of the three axial directions and output signals Ux, Uy, and Uz. As shown in the figure, the axial direction of the rotation shaft 32 of the motor MT is the Z direction, and the X and Y directions are perpendicular to this. Although the acceleration sensors 21-23 are drawn separately in the figure, it is also possible to use a single sensor that can detect accelerations in the three axes and output acceleration signals Ux, Uy, and Uz corresponding to the vibrations of each axis.

[0013] The detection device 10 includes a detection state vector extraction unit 50 that receives the signals Ux, Uy, and Uz output by the acceleration sensors 21-23 and extracts a detection state vector, which is a state vector used for detecting an abnormality; a score generation unit 60 that generates a score for the operating state of the rotor device 30 under test; an identification unit 70 that identifies a failure score for the rotor device 30 from the generated score; a centroid vector storage unit 80 that stores the centroid vector referenced when generating the score; and an output unit 90 that outputs an index corresponding to the likelihood of a failure based on the magnitude of the failure score. The detection device 10 of this embodiment is equipped with an arithmetic and logic operation circuit including a known CPU, GPU, memory, etc., and at least a portion of the detection state vector extraction unit 50, score generation unit 60, etc. are realized by the CPU sequentially executing programs stored in memory. The same applies to the learning mechanism 100. Of course, the detection state vector extraction unit 50, etc. may also be realized using dedicated hardware.

[0014] The sensing state vector extraction unit 50 includes an input layer 51 and a reservoir layer 52. The sensing state vector extraction unit 50 is configured as reservoir computing without an output layer, and the input layer 51 receives signals Ux, Uy, and Uz from the three-axis acceleration sensors 21 to 23. The reservoir layer 52 includes an echo state network, which is a multi-layer neuron network with recurrent connections. The reservoir layer 52 outputs a sensing state vector that represents the state of the rotor device 30.

[0015] The extracted detection state vector is output to the score generation unit 60. The score generation unit 60 calculates and outputs a score representing the state of the rotor device 30 based on the detection state vector and the centroid vector stored in the centroid vector storage unit 80. The identification unit 70 receives the score and identifies a fault score for the rotor device 30. If the fault score is greater than a predetermined value, the identification unit 70 outputs the fault score to the outside via the output unit 90. The generation of a score by the score generation unit 60 requires the centroid vector calculated in advance by the learning mechanism 100 and stored in the centroid vector storage unit 80. Therefore, before describing the configuration of the detection state vector extraction unit 50, the configuration and function of the learning mechanism 100 will be described. The configurations of the learning mechanism 100 and the detection state vector extraction unit 50 are common in that they both utilize the function of reservoir computing. Therefore, describing the learning mechanism 100 will also be helpful in understanding the detection state vector extraction unit 50.

[0016] (A2) Structure and operation of the learning mechanism 100: As already explained, the learning mechanism 100 shown in FIG. 1 is not necessarily required when detecting an abnormality in the operation of the rotor device 30. However, it is necessary to perform learning of the center of gravity vector and the like before the detection device 10 detects an abnormality in the rotor device 30. Therefore, the configuration and operation of the learning mechanism 100 will be explained with reference to FIG. 2. The learning mechanism 100 is configured as a reservoir computing system without an output layer, and includes an input layer 110 that receives signals Ux, Uy, and Uz from the three-axis acceleration sensors 21-23, a reservoir layer 120 that includes an echo state network, which is a multi-layer neuron 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 division unit 140 that divides the numerous stored state vectors into multiple clusters, and a center of gravity vector acquisition unit 150 that calculates the center of gravity vector of each divided cluster. The centroid vector of each cluster obtained by the centroid vector acquisition unit 150 is stored in the centroid vector storage unit 80.

[0017] In this embodiment, the connection weight matrix W 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). The connection weight matrix W in the reservoir layer 120 is set to uniformly distributed random numbers in the interval [-1, 1] (random number seed: 4, connection density: approximately 0.4, ESP requirement: approximately 0.99 / ρ), where ρ is the spectral radius of the matrix.

[0018] When the state vector of the reservoir layer 120 at time t is x(t) and the input data is u(t), the unsupervised learning model (input layer + reservoir layer) shown in the figure is expressed by the following equation (1): x(t-1) represents the state vector at the previous sampling point. x(t)=f(Win·u(t)+W·x(t-1)) …(1) In equation (1), f denotes the activation function, and here tanh is used. The input u(t) in this embodiment is the above-mentioned three-axis acceleration data Ux(t), Uy(t), and Uz(t). Therefore, the number of dimensions of the input is three. When these acceleration data are collectively treated as a vector hereinafter, they are represented as u(t). Furthermore, the dimension of the state vector is set to 64.

[0019] In this embodiment, in order to update the reservoir layer 120 slowly, the state vector x(t) is calculated by the following equation (2) using the leakage rate α. x(t)=(1-α)·x(t-1)+α·f(Win·u(t)+W·x(t-1)) …(2) Here, the leakage rate α was set to a very small value of about 0.0005 to 0.001.

[0020] Figure 3 shows an example of input data u(t) input to an unsupervised learning model consisting of an input layer 110 and a reservoir layer 120. The upper part of the figure shows an example in which continuous rotational vibrations, such as those of a rotor device 30, are used as input data. The lower part shows an example in which data based on a segmented operation, such as that of an injection molding machine, is used, from the start of operation to the completion of product formation. In the case of the rotor device 30, the input data, the three-axis acceleration signals, do not have clear segments. Therefore, the reservoir layer 120 is initialized only once at the beginning. Thereafter, the state vector x(t) is learned at regular intervals based on the input data u(t). Note that, as shown in the figure, the timing of the first learning after initialization is delayed by a predetermined time TRNS from the normal time to avoid the influence of the transient response of the reservoir layer 120 during learning.

[0021] The rotation speed of the rotor device 30 used for learning was 1200 rpm, and the sampling frequency of the three-axis acceleration sensors 21-23 that detect its vibration, i.e., acceleration, was 1 kHz. The input layer 110 downsampled the signals Ux, Uy, and Uz from the acceleration sensors 21-23 to half the frequency, 0.5 kHz, and learned the state vector at 0.5-second intervals, as shown by the timing Xa1, Xa2, etc.

[0022] If the object of learning is an injection molding machine or the like, where the operation is clearly divided into shots, as shown in the bottom part of the figure, the state vector can be initialized INT for each operation cycle, and then the average Xa1, Xa2,... of the state vector during that time can be calculated and learning can be performed.

[0023] When learning the vibration state of the rotor device 30, the length of one learning file was set to approximately 105 seconds. Taking into consideration the transient response period of the acceleration sensor and the uniformity of data length, the first 10 seconds of the file were excluded, leaving 90 seconds of data from 10 seconds to 100 seconds.

[0024] The state vector x(t) thus obtained is stored in the state vector storage unit 130, and the stored state vectors x(t) are divided into a plurality of clusters by the cluster division unit 140. The cluster division may be performed after a large number of state vectors x(t) have been stored, or may be performed sequentially each time a state vector x(t) is obtained, as will be described later.

[0025] The clustering process performed by the cluster division unit 140 will be described in detail later. By clustering, state vectors that are close to each other are grouped. This process is shown schematically in FIG. 4. The state vector x(t) obtained by this embodiment is a multidimensional vector, but for ease of explanation, it is shown as M two-dimensional vectors in the figure. For a set BA of M state vectors x(t), the cluster division unit 140 groups state vectors x(t) that exist in this area BA and are close to each other. In the example shown, the M state vectors x(t) are divided into five clusters L0 to L4. In this case, the centroid vector acquisition unit 150 obtains centroid vectors C0 to C5 for each of the clusters L0 to L4. The obtained centroid vectors are stored in the centroid vector storage unit 80.

[0026] An example of the processing performed by the learning mechanism 100 is shown as a flowchart in FIG. 5. When the learning mechanism 100 starts this processing, it first performs initialization processing (step S201). In the initialization processing, the internal state vectors of the input layer 110 and the reservoir layer 120 are initialized, and as explained with reference to FIG. 3, processing is performed to wait for the passage of a period of transient response of the rotor device 30 to be learned. Next, processing to acquire and save the state vector of the learning object (step S211) is repeated a predetermined number of times (steps STR to STP). When the rotor device 30 is operated and its three-axial accelerations Nx, Ny, and Nz are sampled at a preset sampling rate, processing is performed in the reservoir layer 120 using the sampled time-series data, and the result is acquired as the state vector x(t).

[0027] Once a predetermined number of state vectors x(t) are acquired, next, clustering processing is performed (step S220). Here, sequential clustering processing called SCM was performed. An example of this processing is shown in FIG. 6. In FIG. 5, in order to facilitate correspondence with the block diagram of the learning mechanism 100, it is shown that after M state vectors x(t) are acquired, clustering processing (step S220) is performed. However, since SCM is sequential clustering, it can be executed every time a new state vector x(t) is acquired. After M state vectors x(t) are stored in the state vector storage unit 130, clustering processing other than SCM, for example, K-means++ or hierarchical cluster division method, may be adopted.

[0028] Since the clustering processing by SCM is well-known, detailed description is omitted. After initializing a variable i that specifies which state vector after the processing of the processing target state vector (step S201), the processing shown in FIG. 6 is started. Here, let the i-th state vector be x(i) and the j-th centroid vector be C(j). However, 0≦i<M, 0≦j<N. When a processable state vector x(i) is input (step S221), when the number of clusters is 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, 223). Thereafter, the variable i is incremented (step S224), and until the processing of the M state vectors x(i) is completed (step S225: “yes”), the processing is repeated from step S221 again.

[0029] 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 those already obtained. Subsequently, the pair C(p1) and C(p2) with the minimum distance and its distance dp1-p2 are obtained from the centroid vectors C (steps S222, S231, S232).

[0030] After the above process, the magnitudes of the distances dp0 and dp1-p2 are discriminated. When dp0 < dp1-p2, the centroid vector C(p0) and the input state vector x(i) are integrated, and the value is set to the centroid vector C(p0) (Steps S240, S241). On the other hand, when dp0 ≥ dp1-p2, the centroid vectors C(p1) and C(p2) are integrated, and the value is set to the centroid vector C(p1), and the state vector x(i) is set to the centroid vector C(p2) (Steps S240, S242, 243). Here, the integration of vectors is performed by a weighted weighted average considering the number of state vectors integrated for the centroid vector C. After the above process, again, the increment of the variable i and whether the clustering for the M state vectors is completed are discriminated (Steps S224, 225). If the clustering for the M state vectors is completed (Step S225: "no"), the clustering process is terminated.

[0031] By the clustering process described above, j clusters are formed from the M state vectors, and the centroid vectors C(0) to C(j-1) of the respective clusters are obtained as exemplified in FIG. 4. In terms of FIG. 5, in the sequential clustering process, the calculation of the centroid vector (Step S251) and the storage of the centroid vector (Step S252) are also executed together, and the centroid vector is stored in the centroid vector storage unit 80. If the clustering is not a sequential process, as shown in FIG. 5, the clustering and the calculation / storage of the centroid vector may be performed individually. Above. The learning mechanism 100 described using FIG. 2 and the like is configured separately from the detection device 10 as a detection centroid vector providing device and can also be implemented independently.

[0032] (A3) Abnormality detection of inspection target: The following describes an anomaly detection process performed with the above-described configuration, i.e., the learning mechanism 100 learns the state of the rotor device 30 to be inspected, the M state vectors x(t) are clustered in the centroid vector storage unit 80, and the centroid vector C(j) of each cluster is stored. As shown in FIG. 1 , the detection state vector extraction unit 50 included in the detection device 10 includes an input layer 51 and a reservoir layer 52 having the same configurations as the input layer 110 and the reservoir layer 120 of the learning mechanism 100. Therefore, when signals of the same acceleration are input from the three-axis acceleration sensors 21-23, the detection state vector extraction unit 50 outputs a state vector x(t), similar to the learning mechanism 100. Based on this configuration, the detection device 10 generates a score for the operating state of the rotor device 30 to be inspected, and detects any abnormality in the rotor device 30. FIG. 7 is a flowchart showing a fault score processing routine, which is an example of an anomaly detection process executed by the detection device 10.

[0033] When the detection device 10 starts this fault score processing routine after learning is complete, it acquires a state vector x(t) for detecting the operating state of the rotor device 30 based on signals from the three-axis acceleration sensors 21-23, which are detection time-series data, as the rotor device 30 operates (step S301). The signals from the three-axis acceleration sensors 21-23 are the same as those input to the learning mechanism 100 during learning, but because they are used to detect the operating state of the test object rather than to learn a state vector, they are referred to as "detection time-series data" to distinguish them. Similarly, the state vector extracted by the detection state vector extraction unit 50 has the same configuration as the state vector acquired by the learning mechanism 100 during learning, but to avoid confusion in the explanation, they are referred to as "detection state vectors." When the detection state vector is acquired from the rotor device 30 in operation, the cluster L(j) closest to this detection state vector is identified based on the cluster and centroid vector of the state vector stored in the centroid vector storage unit 80 (step S311). The closest cluster can be identified by calculating the distance from the centroid of each cluster.

[0034] For example, as shown in Figure 8, if the acquired state vector x(t) is indicated by point A, the distance (score) from each of the centroid vectors C(0) to C(4) of each cluster L(0) to L(4) is calculated. Of these scores, the distance from the centroid vector C(3) of the closest cluster L(3) is calculated as the fault score DS (step S321).

[0035] 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, the fault score DS is output as an index representing the degree of the fault (step S341). If the fault score DS is equal to or less than the threshold Tds, no processing is performed. As shown in FIG. 8, the threshold Tds is a predetermined value greater than the average distance from the centroid vector C(j) of each cluster L(j). Therefore, in the illustrated example, if the average of the detected state vector x(i) over a predetermined number of times reaches point A, the fault score DS corresponding to the distance from the centroid vector C(3) of the nearest cluster L(3) will be greater than the threshold Tds. On the other hand, if the average of the detected state vector x(i) over a predetermined number of times reaches point B, the fault score DS corresponding to the distance from the centroid vector C(0) of the nearest cluster L(0) will be less than the threshold Tds. Therefore, in this example, if the average of the state vector x(i) over a predetermined number of times reaches point A, this is determined to be an abnormality, and some kind of abnormality notification is issued from the output unit 90. This abnormality notification is the output of a malfunction indicator in this embodiment. The malfunction indicator may be output by any indicator or method, such as an audio abnormality notification, a display on a printer or screen, or a classification of the detected object (normal product / abnormal product).

[0036] Through the above processing, the anomaly detection system 20 can easily detect an anomaly occurring in the rotor device 30 using a score calculated based on the state vector. Furthermore, when the signals from the three-axis acceleration sensors 21-23 in the rotor device 30 are learned by the learning mechanism 100, which includes the input layer 110 and the reservoir layer 120, there is no need to repeatedly learn the weights in the multi-layer neural network. In this embodiment, since the reservoir computing technique is employed, the connection weight matrix W between the input layer 110 and the reservoir layer 120 and the connection weight matrix W in the reservoir layer 120 can be set to random values in advance, eliminating the need to learn the weights using techniques such as backpropagation. Furthermore, since the state vector in the reservoir layer 120 is used, there is no need to learn the connection weight matrix for the output layer.

[0037] To verify the operation of the anomaly detection system 20 in this embodiment, the difference in score DS was determined under the following conditions. In the rotor device 30, the rotor 33 is disposed eccentrically with respect to the rotation axis 32. Therefore, when the motor MT rotates, vibrations occur throughout the rotor device 30, and the acceleration sensors 21-23 detect the vibrations of the device. The anomaly detection system 20 is required to determine whether this vibration is normal or indicates some kind of abnormality. An abnormality in the rotor device 30 shown in FIG. 1 is thought to occur when the balance of a normal rotor device 30 is disrupted, such as when the first bolt A or the second bolt D is loose. Such imbalances were intentionally created as the following six types of imbalance.

[0038] Type (Code): Contents of imbalance <1> AnDn: Rotor 33 + 1st bolt A + 2nd bolt D <2> AnDw: Rotor 33 + 1st bolt A + 2nd bolt Dw <3> AwDn: Rotor 33 + 1st bolt Aw + 2nd bolt D <4> AwDw: Rotor 33 + 1st bolt Aw + 2nd bolt Dw <5> AnDv: Rotor 33 + 1st bolt A + 2nd bolt Dv <6> AvDn: Rotor 33 + 1st bolt Av + 2nd bolt D In the above, the symbol w indicates that a thin washer is attached to the bolt, the symbol v indicates that a thick washer is attached to the bolt, and the symbol n indicates that no washer is attached to the bolt. 32 files of data were collected for each of <1> to <4>, and 16 files for <5> and <6>. As explained above, each file contains 90 seconds of vibration sampled at a specified interval.

[0039] Figure 9 is a graph showing an example of the average failure score for the above cases <1> to <6>. In the figure, diagonal hatching indicates the graph when clustering was performed using k-means++, and horizontal hatching indicates the graph when clustering was performed using SCM, a sequential clustering method. As shown in the figure, sequential clustering results in a relatively lower average score, but regardless of the clustering method used, the failure score is significantly higher when the first bolt A and the second bolt D are installed improperly (<2> to <6>: with washer) compared to when they are installed properly (<1> AnDn: without washer).

[0040] Figure 10 clearly illustrates this point using the rate of increase in the fault score. In the figure, the broken line JK represents the rate of increase in the fault score when clustering is performed using k-means++, and the broken line JS represents the rate of increase when clustering is performed using SCM, a sequential clustering method. When the first bolt A and the second bolt D are properly installed (〈1〉AnDn: no washer), the fault score is approximately 1.2 to 1.4 times higher when they are not properly installed (〈2〉 to 〈6〉: with washer). Therefore, by appropriately selecting the threshold value Tds, abnormalities in the installation of the rotor device 30 can be easily detected.

[0041] B. Other Embodiments: (1) The present disclosure can also be implemented as a detection device for detecting a driving state of an object to be inspected, as described below. This detection device includes a centroid vector storage unit that stores centroid vectors for each of N (1≦N) clusters obtained by cluster-dividing a plurality of state vectors obtained from time-series data reflecting the operation of the object to be inspected, and a score generation unit that, after the centroid vectors are stored, extracts a detection state vector from detection-time time-series data, which is the time-series data newly obtained regarding the object to be inspected, and generates a score for the driving state of the object to be inspected according to the distance of the detection state vector from the centroid vector.

[0042] This allows us to extract a detection state vector from the newly acquired detection time-series data for the test object and generate a score for the test object's operating condition based on the distance of the detection state vector from the center of gravity vector. For test objects with clearly defined start and end points, such as injection molding machines or assembly robots, time-series data can be acquired for each object. For test objects without clearly defined start and end points, such as motors or engines, acceleration data can be acquired over a certain time period. The acquired time-series data can include data such as acceleration, speed, and position, as well as data representing vibration phenomena (amplitude, frequency, damping rate, resonance, etc.). For electrical devices, electrical data showing changes over time, such as applied voltage and current, can be used. For engines and turbines, time-series data such as pressure and temperature can be used. First- or second-order derivatives of the above data can also be used. Time-series data can be a single piece of data, or multidimensional data, such as triaxial acceleration data. The time series data may also be treated as multidimensional time series data that combines the various types of data described above.

[0043] The operating state of the test object refers to a pre-determined operating state for the test object, such as "normal operation" or "ordinary operation" where the test object is operating without any abnormalities, or "abnormal operation" that deviates from such operating states. The operating state does not necessarily have to be divided into two categories, such as normal (regular) and abnormal, but may be divided into three categories, such as normal (regular), caution required, and abnormal. The behavior of the test object also refers to the movement of the test object itself, such as the rotation of a motor or the molding operation of an injection molding machine. The time-series data reflecting the behavior of the test object need only reflect such behavior of the test object. The test object from which the time-series data referenced when calculating the centroid vector was obtained and the test object from which the detection-time time-series data was obtained when calculating the detection state vector may be the same or different test objects. The centroid vector may be calculated using time-series data obtained from the same model of equipment, and the detection-time time-series data from a different test object may be used to generate the score. For example, if the inspection target is a mass-produced product such as a motor, multiple samples can be used to determine the center of gravity vector, and the specific motor that is actually used can be used to determine the score. Of course, in the case of a large product that is used continuously, such as an injection molding machine, the center of gravity vector can be determined by obtaining time-series data for a specific product, and the detection state vector can be determined from the detection time-series data obtained for the same product.

[0044] The centroid vector storage unit stores the centroid vector of each of N (1≦N) clusters obtained by clustering multiple state vectors obtained from time-series data reflecting the behavior of the test subject. This unit may be prepared before the detection device generates a score, or, if the generated score is determined to be normal when the detection device generates a score, the detection state vector used to generate the score may be used to update the centroid vector of the cluster. Since the score is generated as the distance from the centroid vector of the cluster to the detection state vector, the score may be treated as a scalar quantity, i.e., the distance from one centroid vector, or as a vector of two or more dimensions, i.e., the distance from multiple centroid vectors.

[0045] (2) In the above configuration, the score generation unit may include an identification unit that identifies the smallest distance of the detection state vector from the N centroid vectors as the fault score of the test object. In this way, if the test object is not normal, a fault score reflecting that state can be easily obtained. By generating or identifying the score or fault score, the state of the test object can be traced. For example, if the test object is released to the market and develops an abnormality such as a fault during or after the operational guarantee period, if a score or fault score is obtained, it can be traced to verify the relationship between the initial state and subsequent changes in the test object. Even if the test object does not develop an abnormality such as a fault, the relationship between the score or fault score and the subsequent soundness of the test object can be verified.

[0046] (3) In the configurations (1) and (2) above, an output unit may be provided that outputs an index corresponding to the possibility of failure of the test object when the failure score is equal to or greater than a predetermined value. In this way, it is possible to know the possibility that the test object is in a state that could lead to failure and take appropriate measures, such as correction, repair, reassembly, or sorting before shipping.

[0047] (4) In the configurations (1) to (3) above, the detection device may be provided in a proximity device that is installed close to the test object and directly inputs the time-series data reflecting the test object's movement. The centroid vector storage unit and the score generation unit may be included in the proximity device. This allows immediate reference to the centroid vector, making it easy to generate a score. In the first embodiment described above, the detection device 10 is installed close to the rotor device 30 and is configured as a proximity device that directly inputs the time-series data. In some cases, clustering and calculation of a centroid vector may be performed based on a state vector obtained by inputting the detected time-series data, and the contents stored in the centroid vector storage unit may be updated with the calculated new centroid vector.

[0048] (5) In the configurations (1) to (3) above, the detection device may be provided in proximity to the inspection target, and the score generation unit may be provided in a proximity device that directly inputs the detection time-series data, and the centroid vector storage unit may be provided in a cloud that is connected to the proximity device via a network and is capable of exchanging data with the proximity device. In this way, high-load processing such as the generation of centroid vectors can be performed on the cloud side, simplifying the configuration of the proximity device.

[0049] The configuration of such an anomaly detection system 20A is illustrated in FIG. 11. A detection device 10A, which is a proximity device, is connected to a server 200 located in the cloud via a network NW such as the Internet, and can access a centroid vector storage unit 80A in a storage medium 210, such as a hard disk, within the server 200. The detection device 10A receives signals from the three-axis acceleration sensors 21-23 of the rotor device 30 and outputs a detection state vector using a detection state vector extraction unit 50A. A score generation unit 60A receives this detection state vector and references the centroid vector storage unit 80A of the server 200 via a communication unit 61A and the network NW, and generates a score based on the distance from the centroid vector stored therein. The other configurations are the same as those of the first embodiment.

[0050] (6) In the configurations (1) to (3) above, the detection device may include a proximity device that is installed near the inspection target and that directly inputs the detection time-series data, and the centroid vector storage unit and the score generation unit may be provided in a cloud that is connected to the proximity device via a network and is capable of exchanging data with the proximity device. In this way, high-load processing such as generating centroid vectors and scores can be performed on the cloud side, simplifying the configuration within the proximity device.

[0051] The configuration of such an anomaly detection system 20B is illustrated in FIG. 12. A detection device 10B, which is a nearby device, is connected to a server 200 located in the cloud via a network NW such as the Internet, and is capable of outputting a 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 references a centroid vector storage unit 80B in a storage medium 210 such as a hard disk, and generates a score based on the distance from the centroid vector stored therein. The generated score is sent back to the detection device 10B via the network NW and the communication unit 61B. The other configurations are the same as those in the first embodiment.

[0052] (7) The present disclosure can also be implemented as a detection method for detecting the driving state of an object to be inspected. In this detection method, a plurality of state vectors obtained from time-series data reflecting the operation of the object to be inspected are cluster-divided into N (1≦N) clusters, and the centroid vectors of each of the clusters are stored. After the centroid vectors are stored, a detection state vector is extracted from detection time-series data, which is the time-series data newly obtained for the object to be inspected, and a score of the driving state of the object to be inspected is generated according to the distance of the detection state vector from the centroid vector. According to this detection method, a detection state vector can be extracted from detection time-series data newly obtained for the object to be inspected, and a score of the driving state of the object to be inspected can be generated according to the distance of the detection state vector from the centroid vector. The object to be inspected, etc., are the same as those in the inspection device described above in (1).

[0053] (8) The present disclosure can also be implemented as a device for providing centroid vectors for detection. This device for providing centroid vectors for detection includes a storage unit that stores time-series data reflecting the behavior of an object of inspection as state vectors, a division unit that divides the stored state vectors into N (1≦N) clusters, a centroid vector storage unit that stores centroid vectors calculated for each of the divided clusters, and a providing unit that, in response to a request from an external device, provides the stored centroid vectors for comparison to detect the driving state of the object of inspection. This makes it possible to easily provide an external device that detects the driving state of the object of inspection with centroid vectors calculated for each of the N clusters obtained by dividing the state vectors reflecting the behavior of the object of inspection.

[0054] (9) In the above configuration, the storage unit may include a multi-layer recurrent neuron network to which the time series data is input, and store the output of the neuron network as the state vector. In this way, the time series data can be easily processed and the state vector can be simply stored.

[0055] (10) In the above configuration, the storage unit may operate the neuron network using an ESN (Echo State Network) or LSM (Liquid State Machine) algorithm.

[0056] (11) In each of the above embodiments, some of the configurations realized by hardware may be replaced with software. At least a portion of the configurations realized by software may also be realized by a discrete circuit configuration. Furthermore, when some or all of the functions of the present disclosure are realized by software, the software (computer program) may be provided in a form stored on a computer-readable recording medium. The term "computer-readable recording medium" is not limited to portable recording media such as floppy disks and CD-ROMs, but also includes internal storage devices within a computer, such as various RAMs and ROMs, and external storage devices fixed to a computer, such as a hard disk. In other words, the term "computer-readable recording medium" has a broad meaning, including any recording medium capable of fixing data packets, not just temporarily.

[0057] The present disclosure is not limited to the above-described embodiments and can be realized in various configurations without departing from the spirit thereof. For example, the technical features in the embodiments corresponding to the technical features in each aspect described in the Summary of the Invention section can be appropriately replaced or combined to solve some or all of the above-described problems or achieve some or all of the above-described effects. Furthermore, if a technical feature is not described as essential in this specification, it can be appropriately deleted. [Explanation of symbols]

[0058] 10, 10A, 10B...Detection device, 20, 20A, 20B...Anomaly detection system, 21 to 23...Acceleration sensor, 30...Rotor device, 31...Base, 32...Rotation axis, 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...Identification unit, 80, 80A, 80B...Centroid 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...Centroid vector acquisition unit, 200...Server

Claims

1. A detection device that detects the operating state of an object to be inspected, a centroid vector storage unit that stores centroid vectors for each of N (1≦N) clusters obtained by cluster-dividing a plurality of state vectors obtained from time-series data reflecting the behavior of the test subject; a score generating unit that extracts a detection state vector from detection time-series data, which is the time-series data newly obtained for the test object after the centroid vector is stored, and generates a score for the driving state of the test object according to a distance of the detection state vector from the centroid vector; A detection device comprising:

2. The detection device according to claim 1 , wherein the score generation unit includes an identification unit that identifies the smallest distance among distances of the detection state vector from the N centroid vectors as the failure score of the test object.

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

4. 4. The detection device according to claim 1, wherein the detection device is provided in proximity to the test subject and includes the centroid vector storage unit and the score generation unit within a proximity device that directly inputs the time-series data reflecting the behavior of the test subject.

5. The detection device is The score generation unit is provided in a proximity device that is provided in the vicinity of the test object and that directly inputs the detection time series data, The centroid vector storage unit is provided in a cloud that is connected to the proximity device via a network and is capable of exchanging data with the proximity device. The detection device according to any one of claims 1 to 3.

6. The detection device is a proximity device that is provided in proximity to the inspection object and that directly inputs the detection time series data; a cloud that is connected to the proximity device via a network and is capable of exchanging data with the proximity device, the cloud comprising the centroid vector storage unit and the score generation unit; The detection device according to any one of claims 1 to 3.

7. A detection method for detecting an operating state of an object to be inspected, comprising: a plurality of state vectors obtained from time-series data reflecting the behavior of the test subject are cluster-divided into N (1≦N) clusters, and a center of gravity vector for each of the clusters is stored; After the centroid vector is stored, a detection state vector is extracted from the detection time time series data, which is the newly obtained time series data regarding the test object, and a score of the driving state of the test object is generated according to a distance of the detection state vector from the centroid vector. Detection method.

8. a storage unit that stores time-series data reflecting the behavior of the test object as a state vector; a dividing unit that divides the stored state vectors into N clusters (1≦N); a centroid vector storage unit that stores the centroid vector obtained for each of the divided clusters; a providing unit that provides the stored center of gravity vector for comparison to detect the operating state of the test object in response to an external request; A detection centroid vector providing device comprising:

9. 9. The apparatus for providing a centroid vector for detection according to claim 8, wherein the storage unit includes a multi-layer recurrent neuron network to which the time series data is input, and stores an output of the neuron network as the state vector.

10. The apparatus for providing a centroid vector for detection according to claim 9, wherein the storage unit operates the neuron network using an Echo State Network (ESN) or Liquid State Machine (LSM) algorithm.

Images