Information generating device, determination device, information generating method, determination method, and program

Abstract

An information generation device 1X mainly comprises a feature information acquisition means 14X, a low-variance dimension selection means 15X, and a reference information generation means 16X. The feature information acquisition means 14X acquires feature information of data indicating a normal state. The low-variance dimension selection means 15X selects a low-variance dimension that is a dimension corresponding to a specific number of axes along which the variance of feature information is lowest in a feature space representing the feature information. The reference information generation means 16X generates reference information indicating the normal state in the low-variance dimension on the basis of low-variance dimension feature information representing the feature information in the low-variance dimension.

Description

[Technical field]

[0001] The present disclosure relates to the technical fields of an information generating device, a determination device, an information generating method, a determination method, and a storage medium. [Background technology]

[0002] There are techniques for detecting anomalies by image analysis. For example, Patent Literature 1 discloses a technique for inspecting parts, in which image data representing a normal state is used as training data to train a neural network, and an inspection image is input to the trained neural network to determine whether an anomaly exists in the inspection image. In addition, Non-Patent Literature 1 discloses a technique for determining the presence or absence of an anomaly based on deviation from a feature distribution identified using only normal data in determining an anomaly using an image. Non-Patent Literature 1 also discloses a technique for reducing the feature dimension by principal component analysis. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] International Publication WO2020 / 031984 [Non-patent literature]

[0004] [Non-Patent Document 1] Paolo Napoletano, Flavio Piccoli and Raimondo Schettini, "Anomaly Detection in Nanofibrous Materials by CNN-Based Self-Similarity", MDPI, Sensors 2018, 18, 209. Summary of the Invention [Problem to be solved by the invention]

[0005] In anomaly detection, when the feature dimensions are reduced to reduce processing costs, there is a risk that the accuracy of anomaly detection will also decrease at the same time.

[0006] In consideration of the above-mentioned problems, one of the objectives of the present disclosure is to provide an information generation device, a determination device, an information generation method, a determination method, and a storage medium that realize highly accurate anomaly detection while reducing processing costs, or that generate information necessary to realize such detection. [Means for solving the problem]

[0007] One aspect of the information generating device is A characteristic information acquisition means for acquiring characteristic information of data representing a normal state; a low variance dimension selection means for selecting low variance dimensions which are dimensions corresponding to a predetermined number of axes having the lowest variance of the feature information in a feature space representing the feature information; a reference information generating means for generating reference information representing the normal state in the low variance dimension based on low variance dimension feature information representing the feature information in the low variance dimension; An information generating device having the above configuration.

[0008] One aspect of the information generation method includes: The computer Acquire characteristic information of the data representing a normal state; selecting low-variance dimensions that correspond to a predetermined number of axes having the lowest variance of the feature information in a feature space that represents the feature information; generating reference information representing the normal state in the low variance dimension based on low variance dimension feature information representing the feature information in the low variance dimension; It is a method for generating information.

[0009] program One aspect of the present invention is Acquire characteristic information of the data representing a normal state; selecting low-variance dimensions that correspond to a predetermined number of axes having the lowest variance of the feature information in a feature space that represents the feature information; a program for causing a computer to execute a process of generating reference information representing the normal state in the low variance dimension based on low variance dimension characteristic information representing the characteristic information in the low variance dimension; M It is. Effect of the Invention

[0010] It is possible to achieve highly accurate anomaly detection while reducing processing costs, or to generate information required for achieving this. [Brief description of the drawings]

[0011] [Figure 1] 1 is a schematic configuration of an anomaly detection system according to a first embodiment. [Diagram 2] 1A shows an example of the hardware configuration of an information generating device, and FIG. 1B shows an example of the hardware configuration of a determining device. [Diagram 3] 2 shows an example of a functional block of an information generating device related to information generation processing. [Figure 4] (A) An abstract representation of the distribution of normal points in a two-dimensional feature space. (B) A diagram showing abnormal data and their distances in (A). [Diagram 5] 13 is a graph showing the transition of classification accuracy for multiple sets of dimensionality and distance thresholds. [Figure 6] 11 is a flowchart illustrating an example of a procedure for information generation processing. [Figure 7] 4 is an example of a functional block of a determination device relating to an abnormality determination process. [Figure 8] 13 is a flowchart illustrating an example of a procedure for an abnormality determination process. [Figure 9] 1 shows a schematic configuration of an information generating device according to a second embodiment. [Figure 10] 13 is an example of a flowchart illustrating a processing procedure executed by an information generating device in the second embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] Hereinafter, embodiments of an information generating device, a determination device, an information generating method, a determination method, and a storage medium will be described with reference to the drawings.

[0013] First Embodiment (1) Overall Configuration 1 is a schematic configuration of an anomaly detection system 100 in the first embodiment. The anomaly detection system 100 is a system that detects anomalies in an object photographed by a camera, and executes a process of generating information required for anomaly detection (also referred to as "information generation process") and a process of determining whether an anomaly exists (also referred to as "anomaly determination process"). The anomaly detection system 100 mainly includes an information generation device 1 that performs the information generation process, a storage device 2 that stores information referenced by the information generation device 1 and a determination device 3, and a determination device 3 that performs the anomaly determination process.

[0014] The information generating device 1 executes an information generating process based on the teacher data stored in the teacher data storage unit 21 of the storage device 2, and stores the data generated by the information generating process in the generated data storage unit 22 of the storage device 2.

[0015] The storage device 2 stores information necessary for the information generation process and the anomaly determination process. The storage device 2 may be an external storage device such as a hard disk connected to or built into the information generation device 1 or the determination device 3, may be a storage medium such as a flash memory, or may be a server device that performs data communication with the information generation device 1 and the determination device 3. The storage device 2 may be composed of multiple storage devices and may have the above-mentioned storage units in a distributed manner.

[0016] The storage device 2 functionally includes a teacher data storage unit 21, a generated data storage unit 22, and a judgment target data storage unit .

[0017] The teacher data storage unit 21 stores teacher data, which is image data used in the information generation process by the information generating device 1. The teacher data storage unit 21 includes at least normal teacher data obtained by photographing an object (also simply called an "object") to be subjected to anomaly detection when the object is in a normal state, and abnormal teacher data obtained by photographing the object in an abnormal, non-normal state. Preferably, the normal teacher data is given a label indicating normality (normal label), and the abnormal teacher data is given a label indicating abnormality (abnormal label). Note that the abnormal teacher data is data for verification, which will be described later, and is less in number than the normal teacher data.

[0018] The generated data storage unit 22 stores data generated by the information generation process executed by the information generating device 1. The generated data storage unit 22 mainly stores "dimension reduction information I1", "normal distribution information I2", and "threshold information I3". Here, the dimension reduction information I1 is information (e.g., a transformation matrix) for converting the feature space of the teacher data into a dimension lower than the feature space. As described later, the "dimension lower than the feature space" is a dimension formed by a predetermined number of axes in which the variance of normal teacher data in the feature space is the lowest, and is hereinafter also referred to as a "low variance dimension". The normal distribution information I2 is information representing the distribution (dictionary) of normal teacher data in the low variance dimension. The threshold information I3 represents a threshold for determining whether or not there is an abnormality in the low variance dimension. The details of the method for generating each of these pieces of information will be described later. The normal distribution information I2 and the threshold information I3 correspond to reference information representing the normal state of the object in the low variance dimension.

[0019] The determination target data storage unit 23 stores determination target data used in the abnormality determination process. The determination target data is image data obtained by capturing an image of an object that is to be determined as to whether or not there is an abnormality in the abnormality determination process.

[0020] In addition to the above-mentioned information, the storage device 2 may store various information necessary for the information generation process and the anomaly judgment process. For example, the storage device 2 may store parameters of a feature extractor that extracts features from teacher data or judgment target data. This feature extractor may be obtained by learning any learning model such as deep learning (including convolutional neural network) or support vector machine. For example, when the feature extractor has an architecture based on deep learning, the storage device 2 stores information on various feature extractor parameters such as the layer structure of the feature extractor, the neuron structure of each layer, the number of filters and filter size in each layer, and the weight of each element of each filter.

[0021] The determination device 3 refers to the generated data storage unit 22 and executes anomaly determination processing on the determination target data extracted from the determination target data storage unit 23. When the determination device 3 detects an anomaly based on the determination target data, it may output information on the anomaly (audio output, display, log writing, etc.).

[0022] The configuration of the anomaly detection system 100 shown in FIG. 1 is an example, and various modifications may be made. For example, at least two of the information generating device 1, the storage device 2, and the determination device 3 may be realized by the same device. In another example, the information generating device 1 and the determination device 3 may each be configured by a plurality of devices. In this case, the plurality of devices constituting the information generating device 1 and the plurality of devices constituting the determination device 3 transmit and receive information required to execute a pre-assigned process between the devices by direct wired or wireless communication or communication via a network. Also, instead of handling the determination target data stored in the storage device 2, the determination device 3 may acquire an image generated by a camera that captures an object as the determination target data, and execute an anomaly determination process on the determination target data.

[0023] (2) Hardware Configuration Next, the hardware configurations of the information generating device 1 and the determining device 3 will be described.

[0024] 2(A) shows an example of a hardware configuration of the information generating device 1. The information generating device 1 includes, as hardware, a processor 11, a memory 12, and an interface 13. The processor 11, the memory 12, and the interface 13 are connected via a data bus 19.

[0025] The processor 11 executes a program stored in the memory 12 to function as a controller (arithmetic device) that performs overall control of the information generating device 1. The processor 11 is, for example, a processor such as a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), a TPU (Tensor Processing Unit), or a quantum processor. The processor 11 may be composed of multiple processors. The processor 11 is an example of a computer.

[0026] The memory 12 is composed of various types of volatile and non-volatile memories, such as a random access memory (RAM), a read only memory (ROM), and a flash memory. The memory 12 also stores programs for executing the processes executed by the information generating device 1. Some of the information stored in the memory 12 may be stored in an external storage device, such as the storage device 2, capable of communicating with the information generating device 1, or may be stored in a storage medium that is detachable from the information generating device 1. The memory 12 may also store information stored in the storage device 2 instead.

[0027] The interface 13 is an interface for electrically connecting the information generating device 1 to other devices. These interfaces may be wireless interfaces such as a network adapter for wirelessly transmitting and receiving data to and from other devices, or may be hardware interfaces for connecting to other devices via cables or the like.

[0028] 2B shows an example of a hardware configuration of the determination device 3. The determination device 3 includes, as hardware, a processor 31, a memory 32, and an interface 33. The processor 31, the memory 32, and the interface 33 are connected via a data bus 30.

[0029] The processor 31 executes a program stored in the memory 32 to function as a controller (arithmetic device) that performs overall control of the determination device 3. The processor 31 is, for example, a processor such as a CPU, a GPU, a TPU, or a quantum processor. The processor 31 may be composed of multiple processors. The processor 31 is an example of a computer.

[0030] The memory 32 is composed of various volatile and non-volatile memories such as RAM, ROM, and flash memory. The memory 32 also stores a program for executing the processes executed by the determination device 3. Note that part of the information stored in the memory 32 may be stored in an external storage device such as the storage device 2 capable of communicating with the determination device 3, or may be stored in a storage medium that is detachable from the determination device 3. The memory 32 may also store information stored in the storage device 2 instead.

[0031] The interface 33 is an interface for electrically connecting the determination device 3 to other devices. These interfaces may be wireless interfaces such as a network adapter for wirelessly transmitting and receiving data to and from other devices, or may be hardware interfaces for connecting to other devices via cables or the like.

[0032] The hardware configurations of the information generating device 1 and the determination device 3 are not limited to those shown in Fig. 2(A) and Fig. 2(B). For example, at least one of the information generating device 1 and the determination device 3 may further include a display unit such as a display, an input unit such as a keyboard or a mouse, and a sound output unit such as a speaker.

[0033] (3) Information generation and processing Next, a description will be given of the information generation process executed by the information generating device 1. In summary, the information generating device 1 determines low variance dimensions by selecting axes obtained by applying principal component analysis to feature information of normal teacher data in ascending order of variance, and generates dimension reduction information I1, normal distribution information I2, and threshold information I3. In this way, the information generating device 1 suitably generates information for realizing anomaly determination processing capable of identifying anomalies with low processing cost and high identification performance.

[0034] (3-1) Functional block Fig. 3 is an example of a functional block of the information generating device 1 for information generation processing. As shown in Fig. 3, the processor 11 of the information generating device 1 functionally includes a feature extracting unit 14, a low variance dimension selecting unit 15, and a reference information generating unit 16. Note that in Fig. 3, the blocks through which data is exchanged are connected by solid lines, but the combination of blocks through which data is exchanged is not limited to that shown in Fig. 3. The same applies to other functional block diagrams described later.

[0035] The feature extraction unit 14 extracts teacher data to be used in the information generation process from the teacher data storage unit 21, and performs feature extraction processing on each teacher data. In this case, the feature extraction unit 14 converts each teacher data into feature information (specifically, feature vector) in a feature space with the number of dimensions "N" (N is an integer equal to or greater than 2). In this case, the feature extraction unit 14 configures a feature extractor based on parameters stored in the storage device 2, for example, and inputs each teacher data into the feature extractor in order to obtain a feature vector corresponding to each teacher data. Then, the feature extraction unit 14 supplies the feature vector corresponding to each teacher data to the low variance dimension selection unit 15.

[0036] The low-dispersion dimension selection unit 15 determines N axes (principal component axes) by performing principal component analysis on a plurality of feature vectors corresponding to a plurality of normal teacher data supplied from the feature extraction unit 14, and when arranging the determined N axes in ascending order of dispersion, selects the top "Ns" axes (Ns is an integer of 1 or more satisfying Ns < N) as the axes of the low-dispersion dimension. In this way, the low-dispersion dimension selection unit 15 selects, as the low-dispersion dimension, the dimension corresponding to the Ns axes with the lowest dispersion among the axes obtained by principal component analysis.

[0037] Here, the number of dimensions (i.e., the number of selected axes) Ns may be determined based on verification using abnormal teacher data as described later, or may be determined as an appropriate value pre-stored in the storage device 2 or the like. The low-dispersion dimension selection unit 15 stores information regarding the low-dispersion dimension (for example, a transformation matrix for converting an N-dimensional feature space into an Ns-dimensional low-dispersion dimension) in the generated data storage unit 22 as dimension reduction information I1. Note that the above transformation matrix is information that can be specified from the processing result of principal component analysis. Further, the low-dispersion dimension selection unit 15 converts the feature vector of each teacher data into a vector of the low-dispersion dimension (also referred to as a "low-dispersion dimension feature vector") and supplies the low-dispersion dimension feature vector to the reference information generation unit 16.

[0038] The reference information generation unit 16 generates normal distribution information I2 and threshold information I3 from the low-dispersion dimension feature vectors supplied from the low-dispersion dimension selection unit 15.

[0039] First, the reference information generation unit 16 generates normal distribution information I2 representing the distribution of the low-dispersion dimension feature vectors of the normal teacher data, and stores the generated normal distribution information I2 in the generated data storage unit 22. In this case, for example, the reference information generation unit 16 may calculate the centroid (average) vector, the variance-covariance matrix, etc. of the low-dispersion dimension feature vectors of the normal teacher data as the normal distribution information I2. The normal distribution information I2 is used for calculating the distance described later.

[0040] Furthermore, the reference information generating unit 16 determines the threshold information I3. In this case, for example, the reference information generating unit 16 first calculates the distance (also called "distance L") between the distribution of normal teacher data (for example, the centroid vector of normal teacher data) and the low variance dimension feature vector of each teacher data. Note that the distance L may be any distance used in anomaly detection, such as the Mahalanobis distance using a variance-covariance matrix. Then, the reference information generating unit 16 sets a threshold (also called "distance threshold Lth") for the distance L that can separate the low variance dimension feature vector of the normal teacher data and the low variance dimension feature vector of the abnormal teacher data. In this case, the reference information generating unit 16 sets, for example, a plurality of thresholds that are candidates for the distance threshold Lth prepared in advance, and sets the threshold with the highest discrimination accuracy as the distance threshold Lth. Note that, without being limited to this example, the reference information generating unit 16 may perform a calculation to search for an optimal value of the distance threshold Lth using any numerical analysis method (optimization method). Then, the reference information generating unit 16 stores the threshold information I3 representing the distance threshold Lth in the generated data storage unit 22.

[0041] As described below, the reference information generating unit 16 may determine the distance threshold Lth simultaneously with the number of dimensions Ns through verification using abnormal teacher data. In other words, the low variance dimension selecting unit 15 and the reference information generating unit 16 may simultaneously search for optimal values ​​of the number of dimensions Ns and the distance threshold Lth instead of determining them separately. A specific example of this will be described in detail in the section "(3-3) Determination of the number of dimensions Ns and the distance threshold Lth through verification."

[0042] Each of the components of the processor 11 shown in FIG. 3 (the feature extraction unit 14, the low variance dimension selection unit 15, and the reference information generation unit 16) described above can be realized, for example, by the processor 11 executing a program. Furthermore, each of the components may be realized by recording a necessary program in an arbitrary non-volatile storage medium and installing it as necessary. Note that at least a part of these components may be realized not only by software by a program but also by any combination of hardware, firmware, and software. Furthermore, at least a part of these components may be realized by using an integrated circuit programmable by a user, such as an FPGA (field-programmable gate array) or a microcontroller. In this case, a program composed of the above-mentioned components may be realized by using this integrated circuit. Furthermore, at least a part of each of the components may be configured by an ASSP (Application Specific Standard Produce) or an ASIC (Application Specific Integrated Circuit). In this way, each of the above-mentioned components may be realized by various hardware. Furthermore, each of these components may be realized by the cooperation of multiple computers, for example, using cloud computing technology. The above is the same in other embodiments described later.

[0043] (3-2) Examples of axis selection and their effects Next, a specific example of axis selection of a low variance dimension based on this embodiment and its effects will be described.

[0044] Fig. 4(A) is a diagram that abstractly shows the distribution of feature vectors of normal teacher data (also called "normal data") in a two-dimensional feature space consisting of a first feature axis and a second feature axis. In Fig. 4(A), an ellipse 40 indicates the approximate distribution range of the normal data, and a center of gravity 41 indicates the center of gravity position (position of the center of gravity vector) of the normal data. In Fig. 4(A), the high variance axis a, which is the first principal component axis that maximizes variance, and the low variance axis b, which is the second principal component axis perpendicular to the high variance axis a, are clearly shown by dashed lines.

[0045] In the example of Fig. 4(A), since the low-variance axis b has a lower variance than the high-variance axis a, the low-variance dimension selection unit 15 selects the low-variance axis b as the axis of the low-variance dimension with priority over the high-variance axis a. Here, the low-variance dimension selection unit 15 determines, for example, the one-dimensional low-variance dimension of the low-variance axis b.

[0046] Fig. 4(B) is a diagram showing the feature vector (also called "abnormal data") 42 of the abnormal teacher data in Fig. 4(A) and its distance L. Here, the abnormal teacher data corresponding to the abnormal data 42 has a different distribution in the feature space from the normal teacher data due to being in an abnormal state, and the abnormal data 42 exists outside the ellipse 40 showing the distribution of normal points.

[0047] In this case, as shown in FIG. 4B, the low variance dimension selection unit 15 calculates, as the distance L, for example, the distance (see arrow 43) between the abnormal data 42 and the center of gravity 41 with the low variance axis b as the reference. Here, the abnormal data 42 clearly differs from the distribution of normal points when the low variance axis b is used as the reference. Therefore, the reference information generation unit 16 can appropriately separate the abnormal data 42 from the normal data within the ellipse 40 by setting the distance threshold Lth to the minor axis of the ellipse 40 or a length slightly larger than the minor axis. Note that the calculation of the distance L may take into account the variance-covariance matrix (here, variance) of the normal teacher data, such as the Mahalanobis distance.

[0048] On the other hand, if the high variance axis a is selected as the axis of the low variance dimension to calculate the distance L, the distance L of the abnormal data 42 will have a length equivalent to the arrow 44, and will be shorter than the major axis of the ellipse 40 indicating the distribution of normal points. Therefore, in this case, no matter how the distance threshold Lth is set, it is not possible to set a distance threshold Lth that separates the normal data and the abnormal data 42 within the ellipse 40.

[0049] As shown in the examples of Figures 4(A) and 4(B) described above, an axis with a higher variance represents a feature with a larger variation, and is suitable as an axis representing a feature required for object recognition, but is not an axis representing an effective feature in anomaly detection. On the other hand, an axis with a lower variance represents a feature with a smaller variation, and is suitable as an axis representing a feature common to objects. Therefore, in anomaly detection, a feature dimension that represents a feature common to such objects will be a feature dimension that more clearly shows the difference between the abnormal state and the normal state of the object.

[0050] Considering the above, the information generating device 1 according to the first embodiment preferentially selects an axis with a lower variance as an axis of a lower variance dimension. As a result, the information generating device 1 realizes high accuracy in anomaly detection while realizing high-speed processing by reducing the feature dimension in anomaly detection processing for determining anomalies in unknown data. It becomes.

[0051] (3-3) Determining the number of dimensions Ns and distance threshold Lth through verification Next, a method for determining the number of dimensions Ns and the distance threshold Lth by verification using abnormal teacher data will be described. Here, as an example, a case will be described in which the low variance dimension selection unit 15 and the reference information generation unit 16 cooperate to simultaneously determine the number of dimensions Ns and the distance threshold Lth.

[0052] FIG. 5 is a graph showing the transition of the classification accuracy for a plurality of sets of the number of dimensions Ns and the distance threshold Lth. Here, the graph "G1" shows the transition of the classification accuracy when the number of dimensions Ns is changed while the distance threshold Lth is fixed at "σ", the graph "G2" shows the transition of the classification accuracy when the number of dimensions Ns is changed while the distance threshold Lth is fixed at "2σ", and the graph "G3" shows the transition of the classification accuracy when the number of dimensions Ns is changed while the distance threshold Lth is fixed at "3σ". Here, "σ" indicates the variance of the principal component axis specified by the principal component analysis (for example, the variance of the axis with the lowest variance or the average of the variances of all axes). In addition, the classification accuracy indicates the classification rate for normal teacher data and abnormal teacher data when the corresponding distance threshold Lth and the number of dimensions Ns are provisionally determined.

[0053] As shown in Fig. 5, regardless of the distance threshold Lth employed, the classification accuracy increases as the number of dimensions Ns increases until the number of dimensions Ns reaches a predetermined number (near "Ns5" in Fig. 5). On the other hand, once the number of dimensions Ns reaches the predetermined number or more, the classification accuracy hardly increases even if the number of dimensions Ns increases (i.e., it reaches a plateau).

[0054] Therefore, the low variance dimension selection unit 15 and the reference information generation unit 16 measure the classification accuracy by gradually increasing the number of dimensions Ns for each of a plurality of candidates for the distance threshold Lth, and select the number of dimensions Ns that causes the degree of increase in classification accuracy to be equal to or less than a predetermined degree when the number of dimensions Ns is increased. For example, in the example of Fig. 5, the low variance dimension selection unit 15 and the reference information generation unit 16 select "Ns4" when the distance threshold Lth is σ and 2σ, and select "Ns5" when the distance threshold Lth is 3σ.

[0055] Then, the low variance dimension selection unit 15 and the reference information generation unit 16 adopt the pair with the highest classification accuracy among the multiple pairs of the selected number of dimensions Ns and the corresponding candidates for the distance threshold Lth as the distance threshold Lth and the number of dimensions Ns to be stored in the generated data storage unit 22. In the example of FIG. 5, the low variance dimension selection unit 15 and the reference information generation unit 16 select the distance threshold Lth as 2σ and the number of dimensions Ns as “Ns4” because the pair (2σ, Ns4) has the highest classification accuracy among pairs (σ, Ns4), (2σ, Ns4), and (3σ, Ns5) of the candidates for the distance threshold Lth and the number of dimensions Ns at which the rate of increase in the classification accuracy reaches a plateau. Note that, as described above, the axis of the low variance dimension is selected in the order of decreasing variance, so that the axis selected as the low variance dimension is uniquely determined once the number of dimensions Ns is determined.

[0056] In this way, the low variance dimension selection unit 15 and the reference information generation unit 16 can refer to the classification accuracy calculated by using the abnormal teacher data and determine an appropriate combination of the distance threshold Lth and the number of dimensions Ns. Note that the method of determining the combination of the distance threshold Lth and the number of dimensions Ns is not limited to the method described using FIG. 5, and any optimization method that searches for optimal values ​​of multiple variables may be used.

[0057] (3-4) Processing flow FIG. 6 is an example of a flowchart showing a procedure of the information generating process executed by the information generating device 1.

[0058] First, the information generating device 1 acquires a plurality of pieces of teacher data from the teacher data storage unit 21 (step S11). In this case, for example, the information generating device 1 acquires a plurality of pieces of normal teacher data and abnormal teacher data for verification, which is smaller than the normal teacher data, from the teacher data storage unit 21.

[0059] Next, the information generating device 1 performs a feature extraction process on each of the acquired teacher data (step S12), thereby generating a feature vector represented in a feature space with N dimensions for each of the teacher data.

[0060] Next, the information generating device 1 performs a principal component analysis on the feature vectors of the normal teacher data acquired in step S12 (step S13). Then, the information generating device 1 sorts the principal component axes obtained by the principal component analysis according to the magnitude of variance (step S14).

[0061] Then, the information generating device 1 determines the number of dimensions Ns when selecting an axis from the low variance side and the distance threshold Lth for judging whether it is normal or abnormal based on the distance L (step S15). In this case, for example, the information generating device 1 verifies the classification accuracy using normal teacher data and abnormal teacher data, and determines the number of dimensions Ns and the distance threshold Lth based on the classification accuracy obtained by the verification. Then, the information generating device 1 stores the dimension reduction information I1, normal distribution information I2, and threshold information I3 based on the execution result of step S15 in the generated data storage unit 22 (step S16). In this case, the information generating device 1 stores, for example, the dimension reduction information I1 representing a transformation matrix for transforming from the feature space to the low variance dimension, the normal distribution information I2 representing the distribution (for example, the centroid vector and the variance-covariance matrix, etc.) of the normal teacher data in the low variance dimension, and the threshold information I3 representing the distance threshold Lth in the generated data storage unit 22.

[0062] (4) Abnormality Judgment Processing Next, an explanation will be given of the abnormality determination process executed by the determination device 3. In summary, the determination device 3 converts the feature vector of the determination target data into a low variance dimension feature vector based on the dimension reduction information I1, and determines whether or not the data is abnormal based on the low variance dimension feature vector, the distance L based on the normal distribution information I2, and the distance threshold Lth indicated by the threshold information I3. As a result, the determination device 3 performs anomaly determination of the determination target data accurately and with a low processing load using the low variance dimension in which the axis is selected in ascending order of variance.

[0063] 7 is an example of a functional block diagram of the determination device 3 related to the abnormality determination process. As shown in FIG. 7, the processor 11 of the information generating device 1 functionally includes a feature extractor 34, a low variance dimension converter 35, a comparator 36, and an abnormality determination unit 37.

[0064] The feature extraction unit 34 extracts the data to be determined from the data to be determined storage unit 23, and performs feature extraction processing on the extracted data to be determined. In this case, the feature extraction unit 34 uses the same feature extractor (feature extraction algorithm) as the feature extraction unit 14, and converts the data to be determined into feature information (more specifically, a feature vector) in a feature space with N dimensions.

[0065] The low variance dimension conversion unit 35 converts the feature vector of the data to be determined output by the feature extraction unit 34 into a low variance dimension feature vector, which is a vector in a low variance dimension, based on the dimension reduction information I1 acquired from the generated data storage unit 22. In this case, for example, the low variance dimension conversion unit 35 converts the feature vector of the data to be determined into a low variance dimension feature vector by applying a conversion matrix from the feature space indicated by the dimension reduction information I1 to the low variance dimension.

[0066] The comparison unit 36 ​​calculates the distance L between the distribution in the low variance dimension of the normal teacher data indicated by the normal distribution information I2 acquired from the generated data storage unit 22 and the low variance dimension feature vector of the judgment target data calculated by the low variance dimension conversion unit 35. In this case, the distance L is a distance of the same index as the distance L calculated by the reference information generation unit 16, and may be any distance used in anomaly detection, such as the Mahalanobis distance. For example, the comparison unit 36 ​​calculates the distance L between the centroid vector in the low variance dimension of the normal teacher data indicated by the normal distribution information I2 and the low variance dimension feature vector of the judgment target data.

[0067] The abnormality determination unit 37 determines whether or not there is an abnormality in the data to be determined, based on the distance threshold Lth indicated by the threshold information I3 acquired from the generated data storage unit 22 and the distance L calculated by the comparison unit 36. In this case, for example, when the distance L is equal to or greater than the distance threshold Lth, the abnormality determination unit 37 determines that the data to be determined indicates an object in an abnormal state, and when the distance L is less than the distance threshold Lth, the abnormality determination unit 37 determines that the data to be determined indicates an object in a normal state.

[0068] FIG. 8 is an example of a flowchart showing a procedure of the abnormality determination process executed by the determination device 3.

[0069] First, the feature extraction unit 34 of the determination device 3 acquires the data to be determined from the data storage unit 23 to be determined (step S21). Then, the feature extraction unit 34 of the determination device 3 performs feature extraction processing on the acquired data to be determined (step S22). As a result, the feature extraction unit 34 calculates a feature vector for the data to be determined. Then, the low variance dimension conversion unit 35 of the determination device 3 reduces the dimension of the feature vector obtained in step S22 based on the dimension reduction information I1 acquired from the generated data storage unit 22 (step S23). In this case, the low variance dimension conversion unit 35 converts the feature vector into a low variance dimension feature vector, which is a vector in a low variance dimension.

[0070] Then, the comparison unit 36 ​​of the determination device 3 calculates the distance L for the low variance dimension feature vector of the determination target data obtained in step S23 based on the normal distribution information I2 acquired from the generated data storage unit 22 (step S24). Then, the abnormality determination unit 37 performs abnormality determination of the object indicated by the determination target data based on the distance L calculated in step S24 and the distance threshold Lth indicated by the threshold information I3 acquired from the generated data storage unit 22 (step S25).

[0071] As described above, by using various information generated by the information generating device 1, the judgment device 3 can accurately perform anomaly judgment on the data to be judged in a low-variance dimension suitable for anomaly judgment, while reducing the amount of calculation through dimensionality reduction.

[0072] (5) Variations Next, modified examples that can be applied to the above-described embodiment will be described. The following modified examples may be implemented in any combination.

[0073] (First Modification) The information generating device 1 may execute the information generating process based on normal teacher data without using abnormal teacher data.

[0074] In this case, the normal teacher data is stored in the teacher data storage unit 21. The low variance dimension selection unit 15 selects the axis with the lowest variance by the number of dimensions Ns, which is a fixed value, from each principal component axis obtained by applying the principal component analysis to the feature vector of the normal teacher data calculated by the feature extraction unit 14, and generates dimension reduction information I1. In this case, the number of dimensions Ns, which is a fixed value, is stored in the storage device 2 or the like in advance. The reference information generation unit 16 converts the feature vector of each normal teacher data into a low variance dimension feature vector, and generates normal distribution information I2 based on the low variance dimension feature vector. Furthermore, the reference information generation unit 16 generates threshold information I3 representing the distance threshold Lth based on the variance of the normal teacher data in the principal component axis selected as the axis of the low variance dimension. For example, the reference information generation unit 16 determines the constant multiple of the above-mentioned variance as the distance threshold Lth.

[0075] In this manner, also in this modification, the information generating device 1 can suitably generate various pieces of information (dimensionality reduction information I1, normal distribution information I2, and threshold information I3) necessary for the abnormality determination process based on the normal teacher data.

[0076] (Second Modification) The information generating device 1 selected an axis constituting a low variance dimension from the principal component axes calculated by the principal component analysis. Alternatively, the information generating device 1 may select an axis constituting a low variance dimension without performing the principal component analysis.

[0077] For example, the information generating device 1 may calculate the variance of a feature vector on each axis constituting an N-dimensional feature space based on the feature vector of normal teacher data calculated by the feature extracting unit 14, and select the Nd axes with the lowest variance as axes constituting the low variance dimension. Even in this case, the information generating device 1 can suitably select axes with low variance in a normal state as axes constituting the low variance dimension, and suitably generate various information for performing a highly accurate abnormality determination process.

[0078] <Second embodiment> 9 shows a schematic configuration of an information generating device 1X according to the second embodiment. The information generating device 1X mainly includes a feature information acquiring means 14X, a low variance dimension selecting means 15X, and a reference information generating means 16X. The information generating device 1X may be composed of a plurality of devices. The information generating device 1X may be the information generating device 1 in the first embodiment.

[0079] The feature information acquisition means 14X acquires feature information of data representing a normal state. In this case, the "data" may be image data or sound data. The feature information acquisition means 14X may generate feature information by performing feature extraction processing on the data, or may acquire feature information calculated from the data by another device or processing block, or feature information pre-stored in a storage device. The feature information acquisition means 14X may be, for example, the feature extraction unit 14 in the first embodiment (including modified examples, the same applies below).

[0080] The low variance dimension selection means 15X selects low variance dimensions that are dimensions corresponding to a predetermined number of axes in which the variance of feature information is the lowest in a feature space that represents the feature information. The "axes" may be axes that configure the feature space, or may be axes obtained by applying principal component analysis or the like to the feature space. The low variance dimension selection means 15X may be, for example, the low variance dimension selection unit 15 in the first embodiment.

[0081] The reference information generating means 16X generates reference information representing a normal state in a low variance dimension based on low variance dimension feature information that represents feature information in a low variance dimension. The "low variance dimension feature information" is, for example, the low variance dimension feature vector in the first embodiment. The reference information can be, for example, the normal distribution information I2 and threshold information I3 in the first embodiment. The reference information generating means 16X can be, for example, the reference information generating unit 16 in the first embodiment.

[0082] 10 is an example of a flowchart executed by the information generating device 1X in the second embodiment. First, the feature information acquiring means 14X acquires feature information of data representing a normal state (step S31). Then, the low variance dimension selecting means 15X selects low variance dimensions that are dimensions corresponding to a predetermined number of axes with the lowest variance of feature information in a feature space representing the feature information (step S32). Then, the reference information generating means 16X generates reference information representing a normal state in the low variance dimension based on the low variance dimension feature information that represents the feature information using the low variance dimension (step S33).

[0083] According to the second embodiment, the information generating device 1X can select a dimension in a feature space suitable for anomaly detection, and can also suitably generate reference information that represents a normal state.

[0084] In each of the above-described embodiments, the program can be stored using various types of non-transitory computer readable media and can be supplied to a computer, such as a processor. The non-transitory computer readable medium includes various types of tangible storage media. Examples of the non-transitory computer readable medium include magnetic storage media (e.g., flexible disks, magnetic tapes, hard disk drives), magneto-optical storage media (e.g., magneto-optical disks), CD-ROM (Read Only Memory), CD-R, CD-R / W, and semiconductor memory (e.g., mask ROM, PROM (Programmable ROM), EPROM (Erasable PROM), flash ROM, and RAM (Random Access Memory)). The program may also be supplied to the computer by various types of transitory computer readable media. Examples of the transitory computer readable medium include electric signals, optical signals, and electromagnetic waves. The transitory computer readable medium can supply the program to the computer via a wired communication path such as an electric wire and an optical fiber, or a wireless communication path.

[0085] In addition, some or all of the above embodiments may be described as, but are not limited to, the following supplementary notes.

[0086] [Appendix 1] A characteristic information acquisition means for acquiring characteristic information of data representing a normal state; a low variance dimension selection means for selecting low variance dimensions which are dimensions corresponding to a predetermined number of axes having the lowest variance of the feature information in a feature space representing the feature information; a reference information generating means for generating reference information representing the normal state in the low variance dimension based on low variance dimension feature information representing the feature information in the low variance dimension; An information generating device having the above configuration. [Appendix 2] The information generating device described in Appendix 1, wherein the reference information generating means generates, as the reference information, distribution information representing a distribution of the low variance dimensional feature information and threshold information for determining whether the target data corresponds to the normal state based on a distance between the distribution and the target data. [Appendix 3] The characteristic information acquisition means further acquires abnormality characteristic information which is characteristic information of the data representing the abnormal state, The information generating device according to claim 1 or 2, wherein the low variance dimension selection means determines the number of dimensions of the low variance dimensions based on verification using the abnormality feature information. [Appendix 4] The information generating device described in Appendix 3, wherein the reference information generating means determines, as the reference information, the candidate for reference information corresponding to a pair selected from a plurality of pairs of the candidate for the number of dimensions and the candidate for the reference information based on the verification. [Appendix 5] The information generating device according to any one of appendix 1 to 4, wherein the low variance dimension selection means selects, as the low variance dimensions, dimensions corresponding to a predetermined number of axes selected in order from the side with the lowest variance when axes obtained by performing principal component analysis on the feature information are arranged according to the magnitude of variance of the feature information. [Appendix 6] The information generating device according to any one of appendixes 1 to 5, wherein the low variance dimension selecting means generates conversion information for converting the feature space into the low variance dimension. [Appendix 7] A feature information acquisition means for acquiring feature information of the data to be judged; A low variance dimension conversion means for converting the feature information into low variance dimension feature information represented by a low variance dimension selected by the information generating device according to any one of Supplementary Note 1 to 6; an anomaly determination means for determining the presence or absence of an anomaly in the determination target data based on the low variance dimensional feature information and the reference information generated by the information generating device; A determination device having the above structure. [Appendix 8] The computer Acquire characteristic information of the data representing a normal state; selecting low-variance dimensions that correspond to a predetermined number of axes having the lowest variance of the feature information in a feature space that represents the feature information; generating reference information representing the normal state in the low variance dimension based on low variance dimension feature information representing the feature information in the low variance dimension; Information generation method. [Appendix 9] Acquire characteristic information of the data representing a normal state; selecting low-variance dimensions that correspond to a predetermined number of axes having the lowest variance of the feature information in a feature space that represents the feature information; A storage medium storing a program that causes a computer to execute a process of generating reference information that represents the normal state in the low variance dimension based on low variance dimension characteristic information that represents the characteristic information in the low variance dimension. [Appendix 10] The computer Acquire characteristic information of the data to be judged; The feature information is converted into low variance dimension feature information represented by a low variance dimension selected by the information generating device according to any one of Supplementary Notes 1 to 6; The method includes determining whether or not the target data has an abnormality based on the low variance dimensional feature information and reference information generated by the information generating device. [Appendix 11] Acquire characteristic information of the data to be judged; The feature information is converted into low variance dimension feature information represented by a low variance dimension selected by the information generating device according to any one of Supplementary Notes 1 to 6; A storage medium that stores a program that causes a computer to execute a process of determining whether or not there is an abnormality in the target data based on the low variance dimensional feature information and the reference information generated by the information generating device.

[0087] Although the present invention has been described above with reference to the embodiments, the present invention is not limited to the above-mentioned embodiments. Various modifications that can be understood by a person skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention. In other words, the present invention naturally includes various modifications and corrections that a person skilled in the art could make in accordance with the entire disclosure, including the claims, and the technical ideas. In addition, the disclosures of the above cited patent documents and the like are incorporated herein by reference. [Industrial Applicability]

[0088] The present invention is suitably applied to automatic detection of defective products on product manufacturing lines, nest detection, and any other anomaly detection fields. [Explanation of symbols]

[0089] 1, 1X information generation device 2 Storage device 3 Judgment device 11, 31 processors 12, 32 memory 13, 33 Interface 21 Teacher data storage unit 22 Generated data storage unit 23 Judgment target data storage unit 100 Anomaly Detection System

Claims

1. A characteristic information acquisition means for acquiring characteristic information of data representing a normal state; a low variance dimension selection means for selecting low variance dimensions which are dimensions corresponding to a predetermined number of axes having the lowest variance of the feature information in a feature space representing the feature information; a reference information generating means for generating reference information representing the normal state in the low variance dimension based on low variance dimension feature information representing the feature information in the low variance dimension; An information generating device having the above configuration.

2. The information generating device according to claim 1, wherein the reference information generating means generates, as the reference information, distribution information representing a distribution of the low variance dimensional feature information and threshold information for determining whether the target data corresponds to the normal state based on a distance between the distribution and the target data.

3. The characteristic information acquisition means further acquires abnormality characteristic information which is characteristic information of the data representing the abnormal state, The information generating device according to claim 1 , wherein the low variance dimension selecting means determines the number of dimensions of the low variance dimensions based on verification using the abnormality feature information.

4. The information generating device according to claim 3 , wherein the reference information generating means determines, as the reference information, the candidate for reference information corresponding to a pair selected based on the verification from a plurality of pairs of the candidate for the number of dimensions and the candidate for the reference information.

5. The information generating device according to any one of claims 1 to 4, wherein the low variance dimension selection means selects as the low variance dimensions dimensions corresponding to a predetermined number of axes selected in order from the side with the lowest variance when the axes obtained by performing principal component analysis on the feature information are arranged according to the magnitude of variance of the feature information.

6. A feature information acquisition means for acquiring feature information of the data to be judged; a low variance dimension conversion means for converting the feature information into low variance dimension feature information represented by a low variance dimension selected by the information generating device according to any one of claims 1 to 5; an anomaly determination means for determining the presence or absence of an anomaly in the determination target data based on the low variance dimensional feature information and the reference information generated by the information generating device; A determination device having the above structure.

7. The computer Acquire characteristic information of the data representing a normal state; selecting low-variance dimensions that correspond to a predetermined number of axes having the lowest variance of the feature information in a feature space that represents the feature information; generating reference information representing the normal state in the low variance dimension based on low variance dimension feature information representing the feature information in the low variance dimension; Information generation method.

8. Acquire characteristic information of the data representing a normal state; selecting low-variance dimensions that correspond to a predetermined number of axes having the lowest variance of the feature information in a feature space that represents the feature information; A program that causes a computer to execute a process of generating reference information that represents the normal state in the low variance dimension based on low variance dimension characteristic information that represents the characteristic information in the low variance dimension.

9. The computer Acquire characteristic information of the data to be judged; The feature information is converted into low variance dimension feature information represented by a low variance dimension selected by the information generating device according to any one of claims 1 to 5, The method includes determining whether or not the target data has an abnormality based on the low variance dimensional feature information and reference information generated by the information generating device.

10. Acquire characteristic information of the data to be judged; The feature information is converted into low variance dimension feature information represented by a low variance dimension selected by the information generating device according to any one of claims 1 to 5, A program that causes a computer to execute a process of determining whether or not the target data has an abnormality based on the low variance dimensional feature information and the reference information generated by the information generating device.

Images