Abnormality detection system, learning device, abnormality detection device, learning method, abnormality detection method, and program

Abstract

The present invention comprises a sensor information acquisition unit that acquires sensor information that indicates the state of each of a plurality of monitoring targets as detected by sensors that number fewer than the monitoring targets, a generation unit that generates a classification model that is a one-class support vector machine (SVM) that has been trained on sensor information that corresponds to a normal state for each of the monitoring targets to learn a separation boundary between normal and abnormal for each of the monitoring targets within a feature quantity space that is formed by features of the sensor information in order to classify each of the monitoring targets as normal or abnormal from the relationship between the feature quantities of sensor information to be classified and the separation boundaries, and a classification unit that applies the classification model to sensor information to be classified that has been acquired by the sensor information acquisition unit and thereby classifies each of the monitoring targets as normal or abnormal.

Description

Anomaly detection system, learning apparatus, anomaly detection apparatus, learning method, anomaly detection method, and program

[0001] The present invention relates to an anomaly detection system, a learning apparatus, an anomaly detection apparatus, a learning method, an anomaly detection method, and a program. The present application claims priority based on Japanese Patent Application No. 2025-031708 filed in Japan on February 28, 2025, and Japanese Patent Application No. 2025-139634 filed in Japan on August 25, 2025, the contents of which are incorporated herein by reference.

[0002] There are technologies for detecting anomalies in apparatuses and the like at manufacturing sites and the like. For example, Patent Document 1 discloses a technology for multifaceted analysis of time-series data groups of a predetermined processing unit in a semiconductor manufacturing process by adopting a configuration in which machine learning is performed by inputting time-series data groups measured along with processing of a target object into a plurality of network units for a predetermined processing unit of a manufacturing process. In addition, Non-Patent Document 1 discloses a technology for estimating a kernel function of a quantum feature map and optimizing a conventional SVM.

[0003] Japanese Unexamined Patent Publication No. 2021-086571

[0004] Havlicek V., Corcoles, A.D., Temme, K et al. “Supervised learning with quantum-enhanced feature spaces.” Nature 567, 209-212 (2019). [Searched on February 12, 2025], Internet <URL: https: / / doi.org / 10.1038 / s41586-019-0980-2>

[0005] However, in conventional technologies such as those described in Patent Document 1, one or more sensors are provided for a single object to be monitored, and whether or not an abnormality has occurred in the object to be monitored is detected from the sensor information measured by those one or more sensors. Non-Patent Document 1 also describes how the use of quantum feature spaces by quantum kernels leads to quantum superiority. However, attempts to apply conventional classical feature spaces to anomaly detection in manufacturing sites have been insufficient in terms of the ability to separate and detect abnormalities in multiple objects to be monitored from a small amount of sensor information. As a result, it is necessary to provide as many or more sensors as there are objects to be monitored in order to detect anomalies, and the number of sensors increases in proportion to the number of objects to be monitored, leading to problems such as increased cost, sensor management, and the burden of processing sensor information.

[0006] The present invention has been made in view of the above circumstances, and its objective is to provide an anomaly detection system, learning device, anomaly detection device, learning method, anomaly detection method, and program that can detect whether or not an anomaly has occurred for each of the monitored targets from sensor information measured by a number of sensors smaller than the number of monitored targets.

[0007] The anomaly detection system of the present invention comprises: a sensor information acquisition unit that acquires sensor information by detecting the state of each of a plurality of monitored targets with a number of sensors less than the number of monitored targets; a generation unit that generates a classification model trained to classify whether each monitored target is normal or abnormal based on the relationship between the feature quantities of the sensor information to be classified and the separation boundary in the feature space formed by the features of the learned sensor information, and by learning the separation boundary between normal and abnormal for each monitored target in the feature space formed by the features of the learned sensor information; and a classification unit that classifies whether each monitored target is normal or abnormal by applying the classification model to the sensor information to be classified acquired by the sensor information acquisition unit.

[0008] The learning device of the present invention comprises: a sensor information acquisition unit that acquires sensor information by detecting the state of each of a plurality of monitored targets with a number of sensors less than the number of monitored targets; and a generation unit that is a One-Class SVM (One-Class Support Vector Machine) that has learned the sensor information corresponding to the normal state of each monitored target as unsupervised machine learning, and generates a classification model that has been learned to classify whether each monitored target is normal or abnormal based on the relationship between the features of the sensor information to be classified and the separation boundary in the feature space formed by the features of the learned sensor information.

[0009] The anomaly detection device of the present invention comprises a sensor information acquisition unit that acquires sensor information by detecting the state of each of a plurality of monitored targets with a number of sensors less than the number of monitored targets, and a classification unit that classifies whether each of the monitored targets is normal or abnormal by applying a classification model to the sensor information acquired by the sensor information acquisition unit. The classification model is a One-Class SVM (One-Class Support Vector Machine) that has learned the sensor information corresponding to the normal state of each monitored target, and is a model that has been learned to classify whether each of the monitored targets is normal or abnormal based on the relationship between the feature quantities of the sensor information of the target to be classified and the separation boundary, by learning the separation boundary between normal and abnormal for each monitored target in the feature space formed by the features of the learned sensor information.

[0010] The learning method of the present invention is a learning method performed by a computer used in an anomaly detection system, wherein a sensor information acquisition unit acquires sensor information by detecting the state of each of a plurality of monitored targets with a number of sensors less than the number of monitored targets, and a generation unit is a One-Class SVM (One-Class Support Vector Machine) that has learned the sensor information corresponding to the normal state of each monitored target as unsupervised machine learning, and by learning the separation boundary between normal and abnormal for each monitored target in the feature space formed by the features of the learned sensor information, a classification model is generated that is learned to classify whether each monitored target is normal or abnormal based on the relationship between the features of the sensor information to be classified and the separation boundary.

[0011] The present invention relates to an anomaly detection method performed by a computer used in an anomaly detection system, wherein a sensor information acquisition unit acquires sensor information by detecting the state of each of a plurality of monitored targets with a number of sensors less than the number of monitored targets, and a classification unit classifies each monitored target as normal or abnormal by applying a classification model to the sensor information acquired by the sensor information acquisition unit, wherein the classification model is a One-Class SVM (One-Class Support Vector Machine) that has learned the sensor information corresponding to the normal state of each monitored target, and is a model that has been learned to classify each monitored target as normal or abnormal based on the relationship between the feature quantities of the sensor information of the target to be classified and the separation boundary in the feature space formed by the learned features of the sensor information.

[0012] The present invention is a One-Class Support Vector Machine (SVM) that learns sensor information detected by a computer used in an anomaly detection system, where the state of each of multiple monitored targets is detected by a number of sensors less than the number of monitored targets, and each monitored target corresponds to a normal state. The program learns the separation boundary between normal and abnormal states for each monitored target in the feature space formed by the features of the learned sensor information, thereby generating a classification model trained to classify whether each monitored target is normal or abnormal based on the relationship between the features of the sensor information to be classified and the separation boundary. The program then applies the classification model to the acquired sensor information to be classified, thereby classifying whether each monitored target is normal or abnormal.

[0013] The present invention is a One-Class Support Vector Machine (SVM) that uses unsupervised machine learning to train a computer used in an anomaly detection system. The SVM acquires sensor information from a number of sensors that detect the state of each of multiple monitored targets, fewer than the number of monitored targets. The SVM learns the sensor information corresponding to the normal state of each monitored target. The SVM learns the separation boundary between normal and abnormal states for each monitored target in the feature space formed by the features of the learned sensor information. The SVM then generates a classification model that is trained to classify whether each monitored target is normal or abnormal based on the relationship between the features of the sensor information to be classified and the separation boundary.

[0014] The present invention provides a program that causes a computer used in an anomaly detection system to acquire sensor information by detecting the state of each of multiple monitored objects with a number of sensors less than the number of monitored objects, and to classify each monitored object as normal or abnormal by applying a classification model to the acquired sensor information. The classification model is a One-Class SVM (One-Class Support Vector Machine) that has learned the sensor information corresponding to the normal state of each monitored object, and is a model that has been learned to classify each monitored object as normal or abnormal based on the relationship between the features of the sensor information to be classified and the separation boundary in the feature space formed by the features of the learned sensor information.

[0015] The anomaly detection system of the present invention comprises: a sensor information acquisition unit that acquires sensor information by detecting the state of each of a plurality of monitored targets with a number of sensors less than the number of monitored targets; a preprocessing unit that generates an autoregressive model based on the time-series changes of the sensor information acquired by the sensor information acquisition unit; a generation unit that generates a classification model which is a One-Class SVM (One-Class Support Vector Machine) that has learned the feature quantities of the autoregressive model generated based on the sensor information corresponding to the normal state of each monitored target, and which is trained to classify whether each monitored target is normal or abnormal based on the relationship between the feature quantities of the sensor information of the target to be classified and the separation boundary in the feature space formed by the learned feature quantities of the sensor information; and a classification unit that classifies whether each monitored target is normal or abnormal by applying the classification model to the feature quantities of the autoregressive model generated based on the sensor information of the target to be classified acquired by the sensor information acquisition unit.

[0016] The learning device of the present invention comprises: a sensor information acquisition unit that acquires sensor information by detecting the state of each of a plurality of monitored targets with a number of sensors less than the number of monitored targets; a preprocessing unit that generates an autoregressive model based on the time-series changes of the sensor information acquired by the sensor information acquisition unit; and a generation unit that generates a classification model which is a One-Class SVM (One-Class Support Vector Machine) that has learned the feature quantities of the autoregressive model generated based on the sensor information corresponding to the normal state of each monitored target as an unsupervised machine learning method, and which learns the separation boundary between normal and abnormal for each monitored target in the feature space formed by the learned feature quantities of the sensor information, thereby classifying whether each monitored target is normal or abnormal based on the relationship between the feature quantities of the sensor information to be classified and the separation boundary.

[0017] The anomaly detection device of the present invention comprises: a sensor information acquisition unit that acquires sensor information by detecting the state of each of a plurality of monitored targets with a number of sensors less than the number of monitored targets; a preprocessing unit that generates an autoregressive model based on the time-series changes of the sensor information acquired by the sensor information acquisition unit; and a classification unit that classifies whether each monitored target is normal or abnormal by applying a classification model to the features of the autoregressive model generated based on the sensor information acquired by the sensor information acquisition unit. The classification model is a One-Class SVM (One-Class Support Vector Machine) that has learned the features of the autoregressive model generated based on the sensor information corresponding to the normal state of each monitored target, and is a model that has been learned to classify whether each monitored target is normal or abnormal based on the relationship between the features of the sensor information of the target to be classified and the separation boundary in the feature space formed by the learned sensor information features.

[0018] The learning method of the present invention is a learning method performed by a computer used in an anomaly detection system, wherein a sensor information acquisition unit acquires sensor information by detecting the state of each of a plurality of monitored targets with a number of sensors less than the number of monitored targets, a preprocessing unit generates an autoregressive model based on the time-series changes of the sensor information acquired by the sensor information acquisition unit, and a generation unit is a One-Class SVM (One-Class Support Vector Machine) that has learned the features of the autoregressive model generated based on the sensor information corresponding to the normal state of each monitored target as unsupervised machine learning, and generates a classification model that has been learned to classify whether each monitored target is normal or abnormal based on the relationship between the features of the sensor information of the target to be classified and the separation boundary.

[0019] The present invention relates to an anomaly detection method performed by a computer used in an anomaly detection system, wherein a sensor information acquisition unit acquires sensor information by detecting the state of each of a plurality of monitored targets with a number of sensors less than the number of monitored targets, a preprocessing unit generates an autoregressive model based on the time-series changes of the sensor information acquired by the sensor information acquisition unit, and a classification unit classifies each monitored target as normal or abnormal by applying a classification model to the features of the autoregressive model generated based on the sensor information acquired by the sensor information acquisition unit, the classification model is a One-Class SVM (One-Class Support Vector Machine) that has learned the features of the autoregressive model generated based on the sensor information corresponding to the normal state of each monitored target, and is a model that has been learned to classify each monitored target as normal or abnormal based on the relationship between the features of the sensor information of the target to be classified and the separation boundary in the feature space formed by the learned sensor information features.

[0020] The present invention is a One-Class Support Vector Machine (SVM) that causes a computer used in an anomaly detection system to acquire sensor information by detecting the state of each of multiple monitored targets with a number of sensors less than the number of monitored targets, generates an autoregressive model based on the time-series changes of the acquired sensor information, and learns the features of the autoregressive model generated based on the sensor information corresponding to the normal state of each monitored target. The program then generates a classification model trained to classify whether each monitored target is normal or abnormal based on the relationship between the features of the sensor information of the target to be classified and the separation boundary in the feature space formed by the learned sensor information features, and applies the classification model to the features of the autoregressive model generated based on the acquired sensor information of the target to be classified, thereby classifying whether each monitored target is normal or abnormal.

[0021] The present invention is a One-Class Support Vector Machine (SVM) that, as unsupervised machine learning, learns the features of the autoregressive model generated based on the sensor information corresponding to the normal state of each of the monitored targets, using a computer used in an anomaly detection system. The program learns the separation boundary between normal and abnormal states for each of the monitored targets in the feature space formed by the learned sensor information, thereby generating a classification model that classifies whether each of the monitored targets is normal or abnormal based on the relationship between the sensor information features and the separation boundary.

[0022] The present invention provides a program that causes a computer used in an anomaly detection system to acquire sensor information by detecting the state of each of multiple monitored objects with a number of sensors less than the number of monitored objects, to generate an autoregressive model based on the time-series changes of the acquired sensor information, and to classify each monitored object as normal or abnormal by applying a classification model to the features of the autoregressive model generated based on the acquired sensor information, the classification model is a One-Class SVM (One-Class Support Vector Machine) that has learned the features of the autoregressive model generated based on the sensor information corresponding to the normal state of each monitored object, and is a model that has been learned to classify each monitored object as normal or abnormal based on the relationship between the features of the sensor information of the object to be classified and the separation boundary in the feature space formed by the learned sensor information features.

[0023] According to the present invention, it is possible to detect whether or not an abnormality has occurred for each of the monitored targets based on sensor information measured by a number of sensors smaller than the number of monitored targets.

[0024] This is a diagram showing an example of application of the anomaly detection system 1 in the first embodiment. This is a block diagram showing an example of the configuration of the anomaly detection system 1 in the first embodiment. This is a diagram for explaining the dataset of the first embodiment. This is a diagram for explaining the dataset of the first embodiment. This is a diagram for explaining the dataset of the first embodiment. This is a diagram for explaining the dataset of the first embodiment. This is a diagram for explaining the dataset of the first embodiment. This is a diagram for explaining the processing performed by the preprocessing unit 12 in the first embodiment. This is a diagram for explaining the processing performed by the feature extraction unit 13 in the first embodiment. This is a diagram for explaining the classification model generation processing performed by the generation unit 14 in the first embodiment. This is a diagram for explaining the classification model generation processing performed by the generation unit 14 in the first embodiment. This is a diagram for explaining the classification model generation processing performed by the generation unit 14 in the first embodiment. This is a flowchart showing the processing flow performed by the information processing device 10 in the first embodiment. This is a diagram for explaining the effects of the first embodiment. This is a diagram for explaining the effects of the first embodiment. This is a diagram for explaining the effects of the first embodiment. This is a diagram for explaining the effects of the first embodiment. This is a diagram for explaining the effects of the first embodiment. This is a diagram for explaining the effects of the first embodiment. This is a diagram for explaining the effects of the first embodiment. This is a diagram for explaining the effects of the first embodiment. This is a diagram for explaining the effects of the first embodiment. This is a diagram illustrating the effects of the first embodiment. This is a diagram illustrating the effects of the first embodiment. This is a diagram illustrating an example of application of the anomaly detection system 2 in the embodiment. This is a block diagram illustrating an example of the configuration of the anomaly detection system 2 in the embodiment. This is a diagram illustrating the dataset of the embodiment. This is a diagram illustrating the dataset of the embodiment. This is a diagram illustrating the dataset of the embodiment. This is a diagram illustrating the classification model generation process performed by the generation unit 114 of the embodiment. This is a diagram illustrating the classification model generation process performed by the generation unit 114 of the embodiment. This is a diagram illustrating the classification model generation process performed by the generation unit 114 of the embodiment. This is a diagram illustrating the classification model generation process performed by the generation unit 114 of the embodiment.This is a flowchart illustrating the processing flow performed by the information processing device 100 of the embodiment. This is a diagram illustrating the effects of the embodiment. This is a diagram illustrating the effects of the embodiment. This is a diagram illustrating the effects of the embodiment. This is a diagram illustrating the effects of the embodiment. This is a diagram illustrating the effects of the embodiment. This is a diagram illustrating the effects of the embodiment. This is a diagram illustrating the effects of the embodiment. This is a diagram illustrating the effects of the embodiment. This is a diagram illustrating application examples of the embodiments (first embodiment and second embodiment).

[0025] Hereinafter, a first embodiment of the present invention will be described with reference to the drawings.

[0026] (Regarding the Anomaly Detection System 1) The Anomaly Detection System 1 will be explained using Figures 1 and 2. Figure 1 is a diagram showing an example of the application of the Anomaly Detection System 1 in the first embodiment. Figure 2 is a block diagram showing an example of the configuration of the Anomaly Detection System 1 in the first embodiment.

[0027] As shown in Figure 1, the anomaly detection system 1 is applied to the detection of anomalies in various devices operating in the manufacturing site W. The various devices operating in the manufacturing site W are the monitored objects T (monitored objects T1 to T7). In this figure, monitored object T1 is a belt conveyor. Monitored object T2 is an AGV (Automated Guided Vehicle). Monitored objects T3 to T6 are robot arms. Monitored object T7 is an image inspection device.

[0028] The anomaly detection system 1 collects information indicating the state of the monitored objects T using a number of sensors 20 that is less than the number of monitored objects T. The various pieces of information indicating the state of the monitored objects T can be any physical quantity, such as the rotational speed of power equipment, operating noise, vibration, and temperature. However, it is desirable that the physical quantity be one that can acquire the states of multiple monitored objects T together using a single sensor. For example, a physical quantity such as sound propagating through space can be used as information indicating the state of the monitored objects T.

[0029] It is desirable that the sensor 20 is a non-contact sensor capable of acquiring the state of multiple monitored objects T together with a single sensor. For example, a microphone that collects the sound emitted by each monitored object T can be used as the sensor 20. In the following explanation, the case where the sensor 20 is a microphone will be used as an example, but it is not limited to this. In the combination of a physical quantity indicating the state of the monitored object T and a sensor, any sensor can be used as long as it is capable of acquiring the physical quantities of at least multiple monitored objects T together. In this figure, the sensor 20 is two microphones (sensors 20-1 and 20-2) installed in the manufacturing site W.

[0030] Anomaly detection system 1 is composed of, for example, an information processing device 10, a sensor 20, and a display device 30. The microphone, acting as the sensor 20, collects sounds emitted by each of the multiple monitoring targets T and transmits the collected sound information to the information processing device 10. The sound information here is an example of "sensor information." The information processing device 10 is a computer that identifies whether each of the monitoring targets T is normal or abnormal. The information processing device 10 receives sound information from the sensor 20 and uses the received sound information to identify whether each of the monitoring targets T is normal or abnormal. The information processing device 10 transmits the classification result to the display device 30. The display device 30 is a display that shows images and the like. The display device 30 receives the classification result, which identifies whether each of the monitoring targets T is normal or abnormal, from the information processing device 10 and displays the classification result. The information processing device 10 and the sensor 20, and the information processing device 10 and the display device 30 are connected to each other so as to be able to communicate via a communication network such as the Internet, or via a short-range wireless communication method such as Bluetooth (registered trademark) or infrared communication, or via a wired connection such as a USB cable.

[0031] (Regarding the information processing device 10) As shown in Figure 2, the information processing device 10 includes, for example, a sensor information acquisition unit 11, a preprocessing unit 12, a feature extraction unit 13, a generation unit 14, a classification unit 15, an output unit 16, a dataset storage unit 17, and a classification model storage unit 18.

[0032] The information processing device 10 executes three phases: a training phase, a verification phase, and a classification execution phase. In the training phase, a classification model is generated. The classification model is a model that classifies whether each of the monitored objects T is normal or abnormal based on sensor information. In the verification phase, the classification model generated in the training phase is verified. In the classification execution phase, using the classification model that obtained a predetermined accuracy in the verification phase, anomaly detection is performed to classify whether each of the monitored objects T is normal or abnormal based on the sensor information to be classified.

[0033] The sensor information acquisition unit 11 acquires sensor information and outputs the acquired sensor information to the preprocessing unit 12.

[0034] (Regarding sensor information) In the first embodiment, various datasets are generated using sensor information. A dataset is information in which the sensor information is associated with labels indicating the normal or abnormal state of each monitored object T. The dataset is used in the training phase and the verification phase. Below, an example will be given of generating a dataset that assumes abnormality detection for each of the two monitored objects T, from sound information which contains sounds emitted by each of the two monitored objects T, a belt conveyor CV as monitored object T1, and a vehicle CR (AGV) as monitored object T2. Assuming an abnormality in the belt conveyor CV as monitored object T1, a state in which a foreign object (for example, a nut) is caught between the belt and the pulley is assumed, and the sound emitted by the belt conveyor CV when the pulley is driven with the nut caught is collected as the sound when monitored object T1 is abnormal. Assuming an abnormality in vehicle CR, which is the target of monitoring T2, a bent shaft connecting the left and right tires is assumed. The sound emitted by vehicle CR when the tires are driven with the shaft bent is collected as the sound of an abnormality in target T2.

[0035] Figures 3 (Figures 3A, 3B, 3C), 4, and 5 are diagrams illustrating the dataset of the first embodiment.

[0036] Figure 3A schematically shows a microphone, which functions as a sensor 20, positioned at a first distance (10 cm in this figure) from the conveyor belt CV and a second distance (50 cm in this figure) from the vehicle CR. A dataset is generated using the sound collected by this microphone, which functions as sensor 20. Alternatively, the sounds emitted by two monitored objects T1 and T2 may be collected by the microphone, and the collected sounds may be combined to generate the sound information as a dataset. More specifically, first, first sound information is obtained by collecting the sound emitted by monitored object T1 with the microphone. Next, second sound information is obtained by collecting the sound emitted by monitored object T1 with the microphone. Then, third sound information, which is a combination of the first and second sound information, is used as the dataset. The order in which the sounds are acquired may be reversed. Alternatively, the sounds emitted by monitored object T1 and the sounds emitted by monitored object T2 may be collected using two separate microphones. Figure 3B schematically shows microphone M1 located immediately next to the belt conveyor CV, microphone M2 located 10 cm away, and microphones M3 and M4 located 50 cm away. Figure 3C schematically shows microphone M1 located immediately next to the vehicle CR, microphone M2 located 5 cm away, and microphones M3 and M4 located 50 cm away. By using the sound information collected by each of the microphones M1 to M4 to synthesize the sound emitted by the monitored object T1 and the sound emitted by the monitored object T2, a dataset can be generated. For example, as the sound collected by the sensor 20 shown in Figure 3A, a synthesized sound obtained by synthesizing the sounds collected by microphone M2 in Figure 3B and microphone M4 in Figure 3C can be used. In the first embodiment, each of the two monitored objects T1 and T2 generates four datasets corresponding to the cases (normal / normal, abnormal / abnormal, abnormal / normal, normal / abnormal). (Normal / Normal) corresponds to the case where both monitored object T1 and monitored object T2 are normal. (Abnormal / Abnormal) corresponds to the case where both monitored object T1 and monitored object T2 are abnormal. (Abnormal / Normal) corresponds to the case where monitored object T1 is abnormal and monitored object T2 is normal. (Normal / Abnormal) corresponds to the case where monitored object T1 is normal and monitored object T2 is abnormal.

[0037] Figure 4 schematically illustrates the process of synthesizing the sound of a conveyor belt (CV) and the sound of a vehicle (CR). From left to right, Figure 4 shows information indicating the time-series changes of the sound corresponding to the sound of the conveyor belt (CV), the sound of the vehicle (CR), and the synthesized sound (Mixed data). In the information shown in Figure 4 (information indicating the time-series changes of the sound), the horizontal axis represents time, and the vertical axis represents the loudness of the sound (e.g., amplitude). As shown in this figure, for the sound emitted by the conveyor belt (CV), the start position (Start) and the end position (Stop) of sound generation are identified. Similarly, for the sound emitted by the vehicle (CR), the start position (Start) and the end position (Stop) of sound generation are identified. For example, the start and end positions can be identified based on the amplitude of the sound. Then, the start positions of the two sounds are combined. This makes it possible to generate a synthesized sound in which both sounds are mixed in a balanced manner.

[0038] Figure 5 schematically illustrates the process of extracting information (logging data area) to be used in the dataset from synthesized sound. In the information shown in Figure 5 (information showing the time-series changes of synthesized sound), the horizontal axis represents time, and the vertical axis represents the loudness (e.g., amplitude). As shown in this figure, the dataset is created by deleting the synthesized sound corresponding to the first time (cutting time) after the start of sound generation (Start) and the second time (cutting time) before the end of sound generation (Stop). In other words, the dataset consists of synthesized sound from the point when the first time has elapsed from the start of sound generation (Start) to the second time before the end of sound generation (Stop). This makes it possible to exclude synthesized sound with unstable loudness after the start and before the end of sound generation, and to generate a dataset of synthesized sound with stable loudness.

[0039] (Regarding preprocessing) Returning to the explanation of Figure 1, the preprocessing unit 12 performs preprocessing. Preprocessing here is the process of processing the data (sensor information) so that it is easier for the feature extraction unit 13, which will be described later, to extract features. The preprocessing unit 12 performs preprocessing on the sensor information and outputs the preprocessed sensor information to the feature extraction unit 13.

[0040] The preprocessing unit 12 extracts the loudness of each frequency component contained in the sound information as sensor information as a preprocessing step. For example, the preprocessing unit 12 extracts the loudness of each frequency component by applying a filter that allows a predetermined frequency band to pass through. As the filter group used here, for example, a Mel-Filterbank can be used. A Mel-Filterbank is a group of filters arranged so that the frequency ranges that the filters pass through are equally spaced on the Mel scale. The Mel scale is an acoustic scale based on the human ear's perception of frequency. For example, humans can accurately distinguish between low-frequency sounds (e.g., 500 Hz to 600 Hz), but can hardly distinguish between high-frequency sounds (e.g., 20 kHz to 20.1 kHz). Based on these human audible characteristics, the Mel scale is set so that the frequency band width is narrower for lower frequencies and wider for higher frequencies.

[0041] Figure 6 is a diagram illustrating the processing performed by the preprocessing unit 12 of the first embodiment. Figure 6 schematically shows each waveform obtained by filtering the synthesized sound of the dataset using the filter group in Mel-Filterbank. Each waveform shown in Figure 6 represents the amplitude of the sound component contained in the frequency band corresponding to each of the Mel scales, extracted by applying each of the filters in Mel-Filterbank. In the waveforms shown in Figure 6, the horizontal axis represents time, and the vertical axis represents the amplitude of the sound component. As shown in this figure, the dataset has characteristics in which the magnitudes of the sound components corresponding to each frequency band are different from each other.

[0042] (Regarding feature extraction) Returning to the explanation of Figure 1, the feature extraction unit 13 extracts features from the sensor information. The feature extraction unit 13 extracts features by performing principal component analysis on the sensor information that has been preprocessed by the preprocessing unit 12. The feature extraction unit 13 obtains a predetermined number of principal components through principal component analysis, for example, 10 principal components from the 1st principal component to the 10th principal component. Alternatively, the feature extraction unit 13 sets the number of principal components to be extracted by principal component analysis so that the cumulative contribution exceeds a predetermined ratio (for example, about 0.8).

[0043] FIG. 7 is a diagram for explaining processing performed by the feature quantity extraction unit 13 of the first embodiment. FIG. 7 shows the relationship between the number of principal components and the contribution ratio (the ratio of information held by each element in the total information of sensor information). In FIG. 7, the horizontal axis represents the number of principal components, and the vertical axis represents the contribution ratio. The bar graph in FIG. 7 indicates contribution ratios, and the line graph indicates cumulative contributions. As shown in this figure, the greater the number of principal components, the larger the cumulative contribution. The cumulative contribution increases sharply when the number of principal components is between 1 and 5. When the number of principal components is 5, the cumulative contribution is approximately 0.8. When the number of principal components is between 5 and 10, the cumulative contribution increases gently, and when the number of principal components is 10, the cumulative contribution is approximately 1.0.

[0044] In the training phase, the feature quantity extraction unit 13 outputs feature quantities extracted from sensor information to the generation unit 14. In the classification execution phase, the feature quantity extraction unit 13 outputs feature quantities extracted from sensor information to the classification unit 15.

[0045] (Generation of Classification Model) The generation unit 14 generates a classification model in the training phase. In the training phase, a dataset for learning is used for training to generate the classification model.

[0046] Here, since accurate quality control is performed at the manufacturing site W, it is rare for the monitoring target T to be abnormal, and it is difficult to acquire sensor information corresponding to an abnormal state during normal operation. As a countermeasure against this, in the first embodiment, a classification model is generated by performing unsupervised machine learning. Specifically, as the classification model, a one-class classifier, more specifically, One-Class SVM (One-Class Support Vector Machine) is used. In One-Class SVM, unsupervised learning, that is, only sensor information acquired from the monitoring target T in a normal state is used for learning, and a linear or nonlinear separation boundary formed from a set of sensor information in a normal state is learned. Then, based on the positional relationship between the separation boundary in the feature space and the feature amount of the classification target, an outlier located at a position deviating from the Gaussian distribution (normal distribution) of feature amounts when the feature amount is normal is regarded as abnormal, and classification into normal or abnormal is performed. Therefore, it is not necessary to collect sensor information at the time of abnormality as a data set used for learning. As described above, One-Class SVM is suitable as a method used for abnormality detection of manufacturing equipment and the like in which abnormalities rarely occur.

[0047] Also, in SVM, by using the kernel method, feature amounts extracted from sensor information (training data set) can be mapped to a feature space having a higher dimension than the number of dimensions of the feature amounts, and a nonlinear separation boundary can be expressed from the mapped feature amounts.

[0048] Furthermore, as a kernel method, a more precise separation boundary can be expressed by using a quantum kernel that uses a value obtained by taking the inner product of quantum states for extending the dimension of feature amount information, compared with a conventional kernel as a classical kernel. Therefore, it is possible to classify whether a state is normal or abnormal with higher accuracy using a precise separation boundary.

[0049] The generation unit 14 generates a quantum kernel using, as feature amounts, principal components extracted by the feature amount extraction unit 13 from the training data set (the first to tenth principal components are used). The generation unit 14 generates a classification model by embedding the quantum kernel in an SVM (Support Vector Machine) and causing the SVM to learn the training data set.

[0050] Here, using Figures 8 (Figures 8A and 8B) and 9 (Figures 9A and 9B), the process by which the generation unit 14 generates a classification model will be explained. Figures 8 to 9 are diagrams illustrating the process performed by the generation unit 14 in the first embodiment.

[0051] Figure 8A illustrates the separation process using a classical kernel. Figure 8A also shows an example of the feature distribution mapped to the feature space. The left side of Figure 8A shows an example of a two-dimensional feature space, where the feature axes are represented by two axes, the x and y axes, before dimensional extension. The right side of Figure 8A shows an example of a three-dimensional feature space, where the feature axes are represented by three axes, the x, y, and z axes, after dimensional extension using a classical kernel. In Figure 8A, the RBF (Radial Basis Function) kernel is used as the classical kernel. As shown in this figure, by using a classical kernel to extend the dimension of a two-dimensional feature space, where setting a separation boundary was difficult, to a three-dimensional feature space, it becomes possible to set a separation boundary.

[0052] Figure 8B (reference) illustrates the separation using a quantum kernel. Figure 8B also shows an example of the distribution of features mapped to the feature space. On the left side of Figure 8B, an example of a two-dimensional feature space is shown, where the feature axes are represented by two axes, the x and y axes, as the feature space before dimensional extension. On the right side of Figure 8B, the space on the Bloch sphere (a model that represents the state of a qubit, where 0s and 1s are mixed, as a point on a unit sphere of radius 1) is shown, as the feature space after dimensional extension using a quantum kernel. As shown in this figure, by using a quantum kernel, it is possible to set a separation boundary on the Bloch sphere and form more complex separation boundaries.

[0053] Figure 9A shows an example of a quantum kernel circuit as a basic configuration for generating a quantum kernel QK. A quantum kernel circuit is, for example, a circuit that takes a qubit QBT (qubits) as input and outputs data to a classical bit CBT (bits) after measurement. The initial state of the qubit QBT (qubits) is set to "|0>". In the quantum kernel circuit, the qubit QBT set in the initial state is used to prepare a state vector encoded by the matrix Uk(xi)|0> in the data encoding circuit. To create the inner product, the encoding circuit matrix Uk(xi)† is prepared and the inner product Uk(xi)・Uk(xi)† is calculated. Here, the encoding circuit matrix Uk is a matrix whose elements are the principal components (features) obtained by performing principal component analysis on the k-th sensor information (information after preprocessing). The dagger U† of the encoding circuit matrix is ​​a matrix (dagger matrix) that corresponds to the complex conjugate transpose of matrix U in a unitary matrix. The matrix U(xi) is a matrix whose elements are the principal components (features) xi obtained from certain sensor information. Here, the computational basis is the Z axis, and the projection component in the Z axis direction is measured. The kernel corresponding to the element of each qubit QBT is observed in the subsequent measurement circuit MS, with the computational basis being the Z axis, and its projection component in the Z axis direction is observed. In the measurement circuit MS, the qubit QBT is converted into classical bits CBT (bits) with values ​​of 0 or 1 according to probability.

[0054] Figure 9B shows the circuit configurations of the quantum kernels QK1 and QK2 used by the generation unit 14 in the process of generating the classification model. In this figure, only the configuration of the feature matrix U in the quantum kernel circuit is shown, and the description of the dagger matrix U† corresponding to the Hermitian conjugate is omitted. The quantum kernel circuit for generating quantum kernel QK1 is a circuit in which each of the qubits QBT is provided with an Adamard gate H and a rotation gate Ry with the Y axis as the rotation axis. The quantum kernel circuit for generating quantum kernel QK2 is a circuit in which the least significant bit of the qubit QBT is provided with an Adamard gate H, a CNOT gate (Controlled NOT gete, a gate that connects two qubits), and a rotation gate Rz with the z axis as the rotation axis. Furthermore, the second most significant bit of the qubit QBT is provided with an Adamard gate H, a CNOT gate, and a rotation gate Rz. Furthermore, the circuit is configured such that the most significant bit of the qubit QBT is connected to an Adamard gate H, the least significant bit to a CNOT gate and a rotation gate Ry, and the second most significant bit to be connected to a CNOT gate, a rotation gate Ry, and a rotation gate Rz.

[0055] In the validation phase, the generation unit 14 validates the classification model generated in the training phase. For example, the generation unit 14 validates the model by determining whether it satisfies the generation conditions as a classification model using a validation dataset. The validation dataset consists of four datasets, each corresponding to one of the (normal / normal, abnormal / abnormal, abnormal / normal, normal / abnormal) cases for the two monitored targets T1 and T2. The generation unit 14 can use indicators such as accuracy, recall, and F1 score (F-value) as the generation conditions for the classification model. For example, the generation unit 14 inputs sensor information from the validation dataset into the generated model. The generation unit 14 calculates indicators such as accuracy, recall, and F1 score (F-value) by determining whether the classification result output from the model in response to the input of the validation dataset matches any of the (normal / normal, abnormal / abnormal, abnormal / normal, normal / abnormal) labels assigned to the validation dataset. When the generation unit 14 uses the F1 score as an evaluation metric, it determines that the generation conditions are met if the F1 score exceeds a threshold (for example, from 0.85 to 0.9). On the other hand, if the evaluation metric falls below the threshold, the generation unit 14 determines that the generation conditions are not met. If the model generated in the training phase satisfies the generation conditions as a classification model, the generation unit 14 stores information representing that model (for example, information representing the feature axes that form the feature space and the separation boundary surface) in the classification model storage unit 18 as information representing the classification model. On the other hand, if the model generated in the training phase does not satisfy the generation conditions as a classification model, the generation unit 14 returns to the training phase, generates a new training dataset by increasing the number of datasets, and generates a classification model again using the new training dataset.

[0056] The generator 14 only needs to be able to map the feature quantities to a higher-dimensional feature space (the feature space after mapping) that has more dimensions than the feature space (the feature space before mapping) corresponding to the feature quantities, using the kernel method. The generator 14 may also perform a mapping that extends the dimensions using a classical kernel instead of a quantum kernel.

[0057] (Regarding classification using a classification model) The classification unit 15 uses a classification model to identify whether each of the monitored objects T is normal or abnormal. The classification unit 15 inputs feature quantities of the sensor information to be classified into the classification model that has been verified in the verification phase and determined to satisfy the generation conditions. The feature quantities input into the classification model here are the feature quantities extracted by the feature quantity extraction unit 13 from the sensor information to be classified, which has been preprocessed by the preprocessing unit 12 after the sensor information to be classified has been acquired by the sensor information acquisition unit 11. The classification model classifies whether each of the monitored objects T1 and T2 is normal or abnormal based on the positional relationship between the input feature quantities and the separation boundary surface in the feature quantity space, and outputs the classification result to the output unit 16.

[0058] (Regarding output) The output unit 16 outputs the classification result from the classification unit 15. The output unit 16 outputs the classification result to the display device 30. The display device 30 then displays the classification result. The output unit 16 may also output sensor information acquired by the sensor information acquisition unit 11 and the results of preprocessing performed by the preprocessing unit 12 (for example, the time-series waveform data shown in Figure 6) to the display device 30. The output unit 16 may also output features extracted by the feature extraction unit 13 to the display device 30. In this case, for example, the display device 30 displays the relationship between the first principal component and the second principal component (for example, a diagram as shown in Figures 14A to 14D described later). The dataset storage unit 17 stores the dataset generated by the generation unit 14. The classification model storage unit 18 stores the classification model generated by the generation unit 14.

[0059] The storage unit of the information processing device 10 (including the dataset storage unit 17 and the classification model storage unit 18) is composed of storage media such as HDD (Hard Disk Drive), flash memory, EEPROM (Electrically Erasable Programmable Read Only Memory), RAM (Random Access read / write Memory), ROM (Read Only Memory), or a combination thereof. The storage unit of the information processing device 10 stores programs for executing various processes to realize the functions of the information processing device 10, and temporary data used when performing various processes.

[0060] The functional units of the information processing device 10 (including the sensor information acquisition unit 11, preprocessing unit 12, feature extraction unit 13, generation unit 14, classification unit 15, and output unit 16) are realized by causing the CPU (Central Processing Unit) and / or GPU (Graphics Processing Unit) provided as hardware in the information processing device 10 to execute programs stored in the memory unit of the information processing device 10.

[0061] (Regarding the processing flow performed by the information processing device 10) Here, the processing flow performed by the information processing device 10 will be explained using Figure 10. Figure 10 is a flowchart showing the processing flow performed by the information processing device 10 of the first embodiment.

[0062] First, let's explain the processing flow in the training phase. The sensor information acquisition unit 11 of the information processing device 10 acquires sensor information for training when all monitored objects T are in a normal state (step S10). The preprocessing unit 12 of the information processing device 10 performs preprocessing on the sensor information acquired in step S10 (step S11). The preprocessing unit 12 of the information processing device 10 extracts features by performing principal component analysis on the sensor information preprocessed in step S11 (step S12). The generation unit 14 of the information processing device 10 trains a One-Class SVM model with the features of the training sensor information extracted in step S12 and generates a classification model by training the model using the kernel method (step S13).

[0063] Next, the processing flow in the verification phase will be explained. The sensor information acquisition unit 11 of the information processing device 10 acquires sensor information as information for generating a verification dataset (step S10). The sensor information acquired in the verification phase consists of four sensor data points corresponding to the (normal / normal, abnormal / abnormal, abnormal / normal, normal / abnormal) cases for each of the two monitored targets T1 and T2. By assigning one of the labels (normal / normal, abnormal / abnormal, abnormal / normal, normal / abnormal) to the sensor information acquired here, a verification dataset can be generated. The generation unit 14 of the information processing device 10 verifies the classification model generated in step S13 (step S14). The generation unit 14 inputs the sensor information of the verification dataset into the classification model. The generation unit 14 determines whether the classification result output from the model in response to the input of the verification dataset matches the label of the verification dataset. If the result of classifying each of the verification datasets into the classification model satisfies the generation conditions, the generation unit 14 terminates the generation of the classification model.

[0064] Next, the processing flow in the classification execution phase will be explained. The sensor information acquisition unit 11 of the information processing device 10 acquires sensor information indicating the state of the monitored object T (step S10). The sensor information acquired in the classification execution phase is the sensor information to be classified, and it is unknown whether each of the two monitored objects T1 and T2 corresponds to one of the cases (normal / normal, abnormal / abnormal, abnormal / normal, normal / abnormal). The preprocessing unit 12 of the information processing device 10 performs preprocessing on the sensor information acquired in step S10 (step S11). The preprocessing unit 12 of the information processing device 10 extracts features by performing principal component analysis on the sensor information preprocessed in step S11 (step S12). The classification unit 15 of the information processing device 10 displays a diagram on the display device 30 that maps the features extracted in step S12 to a feature space with the first principal component and the second principal component as feature axes (step S15), and performs classification by the classification model (step S16). The classification model classifies whether each of the monitored objects T1 and T2 is normal or abnormal based on the positional relationship between the feature quantities to be classified and the separation boundary. The classification unit 15 then terminates the processing related to the classification execution.

[0065] Here, the effects of the first embodiment will be explained using Figures 11 to 14. Figures 11 to 14 are diagrams illustrating the effects of the first embodiment.

[0066] Figure 11 shows the relationship between the kernel applied to the classification model and its accuracy. The vertical axis in Figure 11 represents the kernel, where RBF is the classical kernel, QK1 is the quantum kernel QK1 shown in Figure 9B, and QK2 is the quantum kernel QK2 shown in Figure 9B. The horizontal axis in Figure 11 represents the accuracy index, and the vertical axis represents the accuracy. The accuracy shown includes the classification accuracy of the entire model, as well as the classification accuracy of individual cases (normal / normal, abnormal / abnormal, abnormal / normal, normal / abnormal). Ten principal components (features) were used to train the kernel.

[0067] In this figure, the F1 score of the classical kernel was 0.87, the F1 score of the quantum kernel QK1 was 0.97, and the F1 score of the quantum kernel was 1. The F-value F1 is an index calculated by the following formula (1): F1 = 2 × (precision × recall) / (precision + recall) ... (1) where Precision is precision, an index shown as TP / (TP + FP). Recall is recall, an index shown as TP / (TP + FN). TP is true positive, the number of monitored subjects T in a normal state that were correctly classified as "normal". TN is true negative, the number of monitored subjects T in an abnormal state that were correctly classified as "abnormal". FP is false negative, the number of monitored subjects T in a normal state that were incorrectly classified as "abnormal". FN is false positive, the number of monitored subjects T in an abnormal state that were incorrectly classified as "normal".

[0068] The effects of the first embodiment will be explained below by comparing the classical kernel (RBF) and the quantum kernel QK2.

[0069] Figure 12 shows the relationship between the number of principal components (features) and the F1 score as an evaluation metric for the overall classification accuracy of the model. The horizontal axis of Figure 12 represents the number of principal components (features), and the vertical axis represents the F1 score. In this figure, both the classical kernel (RBF) and the quantum kernel QK2 tend to increase in F1 score as the number of principal components (features) increases. The F1 score of the quantum kernel QK2 is approximately 1.0 higher than that of the classical kernel (RBF), indicating that the quantum kernel QK2 has higher accuracy regardless of the number of principal components (features).

[0070] Figures 13 (13A to 13D) show the relationship between the number of principal components (features) and the F1 score as an evaluation index in individual classifications (normal / normal, abnormal / abnormal, abnormal / normal, normal / abnormal). The horizontal axis of Figure 13 shows the number of principal components (features), and the vertical axis shows the F1 score. Figure 13A shows the case where (normal / normal) sensor information is classified into a classification model. In this case, it is shown that both the classical kernel (RBF) and the quantum kernel QK2 have an F1 score of approximately 1.0, regardless of the number of principal components (features). Figure 13B shows the case where (abnormal / abnormal) sensor information is classified into a classification model. In this case, it is shown that both the classical kernel (RBF) and the quantum kernel QK2 have an F1 score of approximately 1.0, regardless of the number of principal components (features). Figure 13C shows the case where (abnormal / normal) sensor information is classified into a classification model. In this case, the F1 score of the classical kernel (RBF) tends to increase slightly as the number of principal components (features) increases, but the F1 score remains around 0.7. The F1 score of the quantum kernel QK2 increases sharply from 0.7 to around 0.95 as the number of principal components (features) increases from 7 to 8, and is shown to reach around 1.0 when the number of principal components (features) is 9 or more. Figure 13D shows a case where (normal / abnormal) sensor information is classified into a classification model. In this case, the F1 score of the classical kernel (RBF) is in the range of approximately 0.8 to 0.86. The F1 score of the quantum kernel QK2 is shown to be in the range of approximately 0.95 to 1.0.

[0071] Figures 14 (14A to 14D) show the relationship between the first and second principal components in the individual classifications (normal / normal, abnormal / abnormal, abnormal / normal, normal / abnormal) of a classification model using the quantum kernel QK2. The horizontal axis of Figure 14 represents the first principal component, and the vertical axis represents the second principal component. Figure 14 verifies whether the abnormalities of two monitored objects T1 and T2 can be classified by mapping the features of sensor information indicating abnormalities in the feature space consisting of the first and second principal components. Figure 14A shows an example where the features of sensor information for (normal / normal) are plotted. In this example, since both monitored object T1 (belt conveyor CV) and monitored object T2 (vehicle CR) are normal, the features of sensor information for (normal / normal) are not plotted in the region where the features of sensor information indicating abnormalities are mapped, and as a result, the features of sensor information for (normal / normal) are not displayed. Figure 14B shows an example where the features of sensor information for (abnormal / abnormal) are mapped. In this example, it is shown that the features are plotted in two distinct regions R1 and R2 in the feature space. This indicates that it is possible to classify two types of anomalies: anomalies in the monitored object T1 (belt conveyor CV) and anomalies in the monitored object T2 (vehicle CR). Figure 14B shows an example of mapping the features of (anomaly / anomaly) sensor information. In this example, it is shown that the features are plotted in two distinct regions R1 and R2 in the feature space. This suggests that the two types of anomalies, anomalies in the monitored object T1 (belt conveyor CV) and anomalies in the monitored object T2 (vehicle CR), can be classified from each other depending on the breakdown of the features obtained by combining the first principal component and the second principal component. Figure 14C shows an example of mapping the features of (anomaly / normal) sensor information. In this example, it is shown that in the feature space, many of the sensor information features when the monitored object T1 (belt conveyor CV) is anomaly are plotted in the same region R1 as shown in Figure 14B. Figure 14D shows an example of mapping the feature quantities of (normal / abnormal) sensor information.In this example, it is shown that in the feature space, many sensor data features are plotted in the same region R2 as shown in Figure 14B, when the monitored object T2 (vehicle CR) is abnormal.

[0072] As described above, the anomaly detection system 1 of the first embodiment comprises a sensor information acquisition unit 11, a generation unit 14, and a classification unit 15. The sensor information acquisition unit 11 acquires sensor information. The sensor information is information obtained by detecting the state of each of a plurality of monitored objects T using a number of sensors 20 that is less than the number of monitored objects T. The generation unit 14 generates a classification model. The classification model is a One-Class SVM (One-Class Support Vector Machine) that has learned the sensor information corresponding to the normal state of each monitored object T. The learned model is a model that learns the separation boundary between normal and abnormal for each monitored object T in the feature space formed by the features of the learned sensor information (sensor information corresponding to the normal state), and is learned to classify whether each monitored object T is normal or abnormal based on the relationship between the feature quantities of the sensor information to be classified and the separation boundary. The classification unit 15 classifies whether each monitored object is normal or abnormal by applying the classification model to the sensor information to be classified. As a result, the anomaly detection system 1 of the first embodiment can generate a classification model that classifies whether each monitored object T is normal or abnormal from sensor information measured by a number of sensors smaller than the number of monitored objects, and can use this classification model to classify whether each monitored object T is normal or abnormal. In other words, it is possible to detect whether or not an anomaly has occurred for each monitored object from sensor information measured by a number of sensors smaller than the number of monitored objects.

[0073] Furthermore, in the anomaly detection system 1 of the first embodiment, the generation unit 14 generates sensor information corresponding to the normal state of each of the multiple monitored targets T by synthesizing the individual detection information (sound information collected by microphones M1 to M4 in Figures 3B and 3C) obtained by each of the sensors detecting the normal state of each of the monitored targets T. The generation unit 14 then trains a One-Class SVM (One-Class Support Vector Machine) with the generated sensor information (sensor information corresponding to the normal state of each of the monitored targets T). As a result, in the anomaly detection system 1 of the first embodiment, even when anomalies are extremely unlikely to occur and it is difficult to acquire sensor information when anomalies occur, it is possible to generate a dataset to be trained on the classification model that performs anomaly detection of the monitored targets T.

[0074] Furthermore, the components of the anomaly detection system 1 of the first embodiment may be divided into a learning device and an anomaly detection device. The learning device is a device that executes the training phase and comprises at least a sensor information acquisition unit 11 and a generation unit 14. The sensor information acquisition unit 11 acquires sensor information for learning. The generation unit 14 generates a classification model. As a result, the learning device of the first embodiment can generate a classification model that classifies whether each of the monitored objects T is normal or abnormal from sensor information measured by a number of sensors smaller than the number of monitored objects. The anomaly detection device is a device that executes the classification execution phase and comprises at least a sensor information acquisition unit 11 and a classification unit 15. The sensor information acquisition unit 11 acquires sensor information to be classified. The classification unit 15 classifies whether each of the monitored objects is normal or abnormal by applying the classification model to the sensor information to be classified. As a result, the anomaly detection device of the first embodiment can classify whether each of the monitored objects T is normal or abnormal using the classification model, from sensor information measured by a number of sensors smaller than the number of monitored objects.

[0075] A second embodiment of the present invention will be described below with reference to the drawings.

[0076] (Regarding the Anomaly Detection System 2) The Anomaly Detection System 2 will be explained using Figures 15 and 16. Figure 15 is a diagram showing an example of the application of the Anomaly Detection System 2 in the second embodiment. Figure 16 is a block diagram showing an example of the configuration of the Anomaly Detection System 2 in the second embodiment.

[0077] As shown in Figure 15, the anomaly detection system 2 is applied to the detection of anomalies in various devices operating in the manufacturing site W. The various devices operating in the manufacturing site W are the monitored objects T (monitored objects T10 to T70). In Figure 15, monitored object T10 is a belt conveyor. Monitored objects T30, T40, T50, and T60 are robot arms. Monitored object T70 is an image inspection device.

[0078] The anomaly detection system 2 collects information indicating the state of the monitored objects T using a number of sensors 200 that is less than the number of monitored objects T. The various pieces of information indicating the state of the monitored objects T can be any physical quantity, such as the rotational speed of power equipment, operating noise, vibration, and temperature. However, it is desirable that the physical quantity be one that can acquire the states of multiple monitored objects T together using a single sensor. For example, a physical quantity such as sound propagating through space can be used as information indicating the state of the monitored objects T.

[0079] It is desirable that the sensor 200 is a non-contact sensor capable of acquiring the state of multiple monitored objects T together with a single sensor. For example, a microphone that collects the sound emitted by each of the monitored objects T can be used as the sensor 200. In the following explanation, the case where the sensor 200 is a microphone will be used as an example, but it is not limited to this. In the combination of a physical quantity indicating the state of the monitored objects T and a sensor, any sensor can be used as long as it is capable of acquiring the physical quantities of at least multiple monitored objects T together. In Figure 15, the sensor 200 is two microphones (sensors 200-1 and 20-2) installed in the manufacturing site W.

[0080] Anomaly detection system 2 is composed of, for example, an information processing device 100, a sensor 200, and a display device 300. The microphone, which acts as the sensor 200, collects sounds emitted by each of the multiple monitored objects T and transmits the collected sound information to the information processing device 100. The sound information here is an example of "sensor information." The information processing device 100 is a computer that identifies whether each of the monitored objects T is normal or abnormal. The information processing device 100 receives sound information from the sensor 200 and uses the received sound information to identify whether each of the monitored objects T is normal or abnormal. The information processing device 100 transmits the classification result to the display device 300. The display device 300 is a display that shows images and the like. The display device 300 receives the classification result, which identifies whether each of the monitored objects T is normal or abnormal, from the information processing device 100 and displays the classification result. The information processing device 100 and the sensor 200, and the information processing device 100 and the display device 300 are connected to each other via a communication network such as the Internet, or via a short-range wireless communication method such as Bluetooth (registered trademark) or infrared communication, or via a wired connection such as a USB cable.

[0081] (Regarding the information processing device 100) As shown in Figure 16, the information processing device 100 includes, for example, a sensor information acquisition unit 111, a preprocessing unit 112, a feature extraction unit 113, a generation unit 114, a classification unit 115, an output unit 116, a dataset storage unit 117, and a classification model storage unit 118.

[0082] The information processing device 100 executes three phases: a training phase, a verification phase, and a classification execution phase. In the training phase, a classification model is generated. The classification model is a model that classifies whether each of the monitored objects T is normal or abnormal based on sensor information. In the verification phase, the classification model generated in the training phase is verified. In the classification execution phase, using the classification model that obtained a predetermined accuracy in the verification phase, anomaly detection is performed to classify whether each of the monitored objects T is normal or abnormal based on the sensor information to be classified.

[0083] The sensor information acquisition unit 111 acquires sensor information and outputs the acquired sensor information to the preprocessing unit 112.

[0084] (Regarding sensor information) In the second embodiment, various datasets are generated using sensor information. A dataset is information in which labels indicating the normal or abnormal state of each monitored object T are associated with the sensor information. The dataset is used in the training phase and the verification phase. Below, an example will be given of generating a dataset that assumes abnormality detection for each of the two monitored objects T, using sound information that contains sounds emitted by each of the two monitored objects T, a conveyor CON as monitored object T10, and a chain belt CHA as monitored object T2. In the case of the conveyor CON as monitored object T10, the belt runs by the drive of a pulley while in contact with a metal frame, and the sound emitted by the conveyor CON when the pulley is driven with a foreign object (for example, a nut or nail) caught between the belt and the metal frame is collected as the sound of an abnormality in monitored object T10. As an abnormality in the chain belt CHA, which is the monitored object T2, a pointed foreign object (for example, a nut or nail) is positioned so that its tip touches the chain belt in the direction of rotation. When the foreign object swings like a pendulum and touches or separates from the chain belt, the sound emitted by the chain belt CHA is collected by the monitored object T2 as the sound of the abnormality.

[0085] Figures 17 and 18 (Figures 18A, 18B, and 18C) are diagrams illustrating the dataset of the second embodiment.

[0086] Figure 17 shows the breakdown of the datasets used in the training and verification phases of the second embodiment. In Figure 17, (0) indicates that the monitored object T is normal, and (1) indicates that the monitored object T is abnormal. More specifically, in the combination of conveyor CON / chain belt CHA, (0 / 0) indicates that both conveyor CON and chain belt CHA are normal. (0 / 1) indicates that both conveyor CON and chain belt CHA are abnormal. (1 / 0) indicates that both conveyor CON and chain belt CHA are abnormal. (1 / 1) indicates that both conveyor CON and chain belt CHA are abnormal. In Figure 17, it is shown that the total dataset used consisted of 60 normal / normal (0 / 0) datasets and 30 each of normal / abnormal (0 / 1), abnormal / normal (1 / 0), and abnormal / abnormal (1 / 1) datasets. Furthermore, the breakdown of the datasets shows that of the 60 normal / normal (0 / 0) datasets, 40 were used in the training phase, and the remaining 20 were used in the testing phase. In addition, 30 datasets each of normal / abnormal (0 / 1), abnormal / normal (1 / 0), and abnormal / abnormal (1 / 1) were used in the testing phase.

[0087] Figure 18A shows an example of a plan view of the measurement environment. This measurement environment is for acquiring data sets and simulates a manufacturing site where multiple monitoring targets T are installed and operating. In Figure 18A, the measurement environment is equipped with a conveyor CON, a chain belt CHA, and a rubber belt machine Rub. The conveyor CON is installed to collect normal sounds emitted when monitoring target T10 is operating normally in the manufacturing site, and abnormal sounds emitted when an abnormality occurs. The chain belt CHA is installed to collect normal sounds emitted when monitoring target T2 is operating normally in the manufacturing site, and abnormal sounds emitted when an abnormality occurs. The rubber belt machine Rub is installed to collect noise (white noise) generated in the manufacturing site.

[0088] The conveyor CON, chain belt CHA, and rubber belt machine Rub are installed facing each other along the y-axis (y-axis). The distance from each of the conveyor CON, chain belt CHA, and rubber belt machine Rub to the x-axis (x-axis) is 0.4 m. Also, in Figure 18A, four microphones M (microphones M10, M20, N30, M40) are installed spaced apart from each other along the x-axis. More specifically, microphone M10 is installed at the origin, microphone M20 is installed 1 m away from the installation position of microphone M10 in the x-axis direction, microphone M30 is installed 2 m away from the installation position of microphone M10 in the x-axis direction, and microphone M40 is installed 3 m away from the installation position of microphone M10 in the x-axis direction.

[0089] Each of the four microphones M (microphones M10, M20, N30, and M40) is a directional microphone with three channels CH (channels CH1 to CH3) each having a directional direction set in three directions at 45-degree intervals in a 360-degree direction. Each of the four microphones M (microphones M10, M20, N30, and M40) collects sound (normal sound or abnormal sound, and noise) arriving at the installed location through its respective channel CH (channels CH1 to CH3).

[0090] Figure 18B shows the sound sources (monitored objects T) that each of the four microphones M (microphones M10, M20, N30, and M40) was primarily able to collect on its respective channel CH (channels CH1 to CH3). In Figure 18B, it is shown that microphone M10's channel CH1 was able to collect sound emitted from the chain belt CHA. It is shown that microphone M10's channel CH2 was able to collect sound emitted from both the conveyor CON and the chain belt CHA. It is shown that microphone M10's channel CH3 was primarily able to collect sound emitted from the conveyor CON. Furthermore, it is shown that microphones M20, M30, and M40's channels CH1 to CH3 were able to collect sound emitted from both the conveyor CON and the chain belt CHA.

[0091] Figure 18C shows an example of a side view of a conveyor CON and a chain belt CHA installed in the measurement environment. As shown in Figure 18C, when collecting the normal sound of the chain belt CHA, the chain belt CHA is operated without any foreign object N3 present. When collecting the normal sound of the conveyor CON, the conveyor CON is operated without the two foreign objects N1 and N2 being caught between them. On the other hand, when collecting the abnormal sound of the chain belt CHA, the foreign object N3 is positioned on the chain belt CHA so as to strike it in the direction of the chain belt's rotation. Also, when collecting the abnormal sound of the conveyor CON, the two foreign objects N1 and N2 are caught between the belt and the metal plate of the conveyor CON.

[0092] (Regarding preprocessing) Returning to the explanation of Figure 15, the preprocessing unit 112 performs preprocessing. Preprocessing here is the process of processing the data (sensor information) so that it is easier for the feature extraction unit 113, which will be described later, to extract features. The preprocessing unit 112 performs preprocessing on the sensor information and outputs the preprocessed sensor information to the feature extraction unit 113.

[0093] The preprocessing unit 112 models the time-series data of the sensor information as a preprocessing step. Methods for modeling time-series data include, for example, AR (AutoRegressive), MA (Moving Average), ARMA (AutoRegressive Moving Average), ARIMA (AutoRegressive Integrated Moving Average), and SARIMA (Seasonal ARIMA, which considers seasonal variations). The following description will illustrate the case where the preprocessing unit 112 generates an AR model to fit the time-series data of the sensor information.

[0094] The preprocessor 112 can estimate the parameters of the AR model using, for example, the Yule-Waler method and the Levinson-Durbin recursive algorithm, which solves the Yule-Walker equations quickly.

[0095] The AR model can be represented by the following equation (1).

[0096]

[0097] Here, in equation (1), yt is the sensor information value at time t. c is a constant. p is the order. φi is the coefficient in order i. yT10 is the sensor information value at time T10. εt is the noise at time t. The noise εt is proportional to the white noise (W.N.) with distance σ as the variable. Distance σ is the distance corresponding to the x-coordinate of the microphone M placed on the x-axis in Figure 18A. Noise emitted from the rubber belt machine Rub installed in the measurement environment can be used as white noise.

[0098] (Regarding feature extraction) Returning to the explanation of Figure 15, the feature extraction unit 113 extracts features from the sensor information. The feature extraction unit 113 extracts the coefficients φi of the AR model generated by the preprocessing unit 112 as features. Here, the subscript i of the coefficient φ is the order (model order) shown in equation (1). The subscript i can take values ​​from 1 to p, where p is any integer greater than or equal to 1.

[0099] Here, the feature extraction unit 113 extracts the coefficients φ1, φ2, φ3, φ4, and φ5 of the AR model shown in equation (1) as features. This is because, based on preliminary experiments verified with various model orders using the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), which are indices for evaluating the model, it was determined that p=5 provides the optimal balance between model complexity and the degree of fit to sensor information (sound data). However, it is not limited to this, and coefficients φi (where the subscript i is a value from 1 to p) corresponding to any value p can be extracted as features.

[0100] During the training phase, the feature extraction unit 113 outputs the features extracted from the AR model to the generation unit 114. During the classification execution phase, the feature extraction unit 113 outputs the features extracted from the AR model to the classification unit 115.

[0101] (Regarding the generation of the classification model) The generation unit 114 generates a classification model in the training phase. In the training phase, the training dataset is used for training to generate the classification model.

[0102] Here, because accurate quality control is carried out at the manufacturing site W, it is rare for the monitored object T to be abnormal, and it is difficult to obtain sensor information corresponding to abnormalities during normal operation. As a countermeasure, in the second embodiment, a classification model is generated by performing unsupervised machine learning. Specifically, a one-class classifier, more specifically a One-Class SVM (One-Class Support Vector Machine), is used as the classification model. In One-Class SVM, unsupervised learning is used, meaning that only features obtained from the monitored object T in a normal state (coefficients of the AR model generated from sensor information) are used for learning, and a linear or nonlinear separation boundary composed of normal features is learned. Based on the positional relationship between the separation boundary in the feature space and the features to be classified, outliers where the features are located outside the Gaussian distribution (normal distribution) of the features when normal are considered abnormal, and the classification of whether it is normal or abnormal is performed. For this reason, it is not necessary to collect sensor information when abnormal as a dataset to be used for learning. Thus, One-Class SVM is a suitable method for detecting anomalies in manufacturing equipment and other devices where malfunctions rarely occur.

[0103] Furthermore, in SVM, by using the kernel method, the coefficients of the AR model generated from sensor information can be mapped as features to a feature space with a higher dimension than the number of dimensions of those features, and a nonlinear separation boundary can be represented from the mapped features.

[0104] Furthermore, as a kernel method, by using a quantum kernel, which utilizes the inner product of quantum states to extend the dimension of feature information, in addition to conventional classical kernels, a more precise separation boundary can be represented. Therefore, using a precise separation boundary, it is possible to classify whether something is normal or abnormal with higher accuracy.

[0105] The generation unit 114 generates a quantum kernel using the coefficients φi (coefficients φ1, φ2, φ3, φ4, φ5) extracted from the AR model extracted by the feature extraction unit 113 for the training dataset as features. The generation unit 114 embeds the quantum kernel into an SVM (Support Vector Machine) and generates a classification model by training it with the training dataset.

[0106] Here, the process by which the generation unit 114 generates a classification model will be explained using Figures 19 (Figures 19A and 19B) and 20 (Figures 20A and 20B). Figures 19 and 20 are diagrams illustrating the process performed by the generation unit 114 in the second embodiment.

[0107] Figure 19A illustrates the separation process using a classical kernel. Figure 19A also shows an example of the distribution of features mapped to the feature space. The left side of Figure 19A shows an example of a two-dimensional feature space, where the feature axes are represented by two axes, the x and y axes, before dimensional extension. The right side of Figure 19A shows an example of a three-dimensional feature space, where the feature axes are represented by three axes, the x, y, and z axes, after dimensional extension using a classical kernel. In Figure 19A, the RBF (Radial Basis Function) kernel is used as the classical kernel. As shown in Figure 19A, by using a classical kernel, it becomes possible to establish a separation boundary in a two-dimensional feature space, where setting a separation boundary was previously difficult, by extending the dimension to a three-dimensional feature space.

[0108] Figure 19B illustrates the separation using a quantum kernel. Figure 19B also shows an example of the distribution of features mapped to the feature space. The left side of Figure 19B shows an example of a two-dimensional feature space, where the feature axes are represented by two axes, the x and y axes, before dimensional extension. The right side of Figure 19B shows the feature space after dimensional extension using a quantum kernel, specifically the space on a Bloch sphere (a model representing the state of a qubit, a mixture of 0s and 1s, as a point on a unit sphere of radius 1). As shown in Figure 19B, by using a quantum kernel, a separation boundary can be established on the Bloch sphere, making it possible to form more complex separation boundaries.

[0109] Figure 20A shows an example of a quantum kernel circuit as a basic configuration for generating a quantum kernel QK. A quantum kernel circuit is, for example, a circuit that takes a qubit QBT (qubits) as input and outputs data to a classical bit CBT (bits) after measurement. The initial state of the qubit QBT (qubits) is set to "|0>". In the quantum kernel circuit, the qubit QBT set to the initial state is encoded by the data encoding circuit matrix Φ(xi)|0> in the data encoding circuit to prepare a state vector. To create the inner product, the encoding circuit matrix Φ(xi)† is prepared and the inner product Φ(xi)・Φ(xj)† is calculated. Here, the encoding circuit matrix Φ(xi) is a matrix whose element is xi, which is the i-th feature (the coefficient of dimension i in the AR model generated from sensor information). The dagger Φ† of the encoding circuit matrix is ​​a matrix (dagger matrix) that corresponds to the complex conjugate transpose of the matrix Φ in the unitary matrix. Here, the computational basis is defined as the Z-axis, and the projection component in the Z-axis direction is measured. The kernel corresponding to each element of the qubit QBT is observed in the subsequent measurement circuit MS, with the computational basis as the Z-axis, and its projection component in the Z-axis direction is observed. In the measurement circuit MS, the qubit QBT is converted into classical bits CBT (bits) with values ​​of 0 or 1 according to probability.

[0110] Figure 20B shows the circuit configuration of the encoding circuit matrix Φ(xi) used by the generation unit 114 in the process of generating the classification model. In Figure 20B, only the configuration of the feature matrix Φ in the quantum kernel circuit is shown, and the description of the dagger matrix Φ† corresponding to the Hermitian conjugate is omitted. The encoding circuit matrix Φ(xi) is a circuit in which each qubit QBT is provided with a Hadamard gate H, a CNOT gate (Controlled NOT gete, a gate that connects two qubits), and a rotation gate Ry with the Y axis as the rotation axis.

[0111] In the validation phase, the generation unit 114 validates the classification model generated in the training phase. For example, the generation unit 114 validates the model by determining whether it satisfies the generation conditions as a classification model using a validation dataset. The validation dataset uses four datasets, each corresponding to one of the cases (normal / normal, abnormal / abnormal, abnormal / normal, normal / abnormal) for the two monitored targets T10 and T2. The generation unit 114 can use indicators such as accuracy, recall, and F1 score (F-value) as the generation conditions for the classification model. For example, the generation unit 114 inputs the features of the validation dataset (coefficients of the AR model generated from sensor information in the validation dataset) into the generated model. The generation unit 114 calculates indicators such as accuracy, recall, and F1 score (F-value) by determining whether the classification result output from the model in response to the input of the features of the validation dataset matches any of the labels (normal / normal, abnormal / abnormal, abnormal / normal, normal / abnormal) assigned to the validation dataset. When the generation unit 114 uses the F1 score as an evaluation metric, it determines that the generation conditions are met if the F1 score exceeds a threshold (for example, from 0.85 to 0.9). On the other hand, if the evaluation metric falls below the threshold, the generation unit 114 determines that the generation conditions are not met. If the model generated in the training phase satisfies the generation conditions as a classification model, the generation unit 114 stores information representing that model (for example, information representing the feature axes that form the feature space and the separation boundary surface) in the classification model storage unit 118 as information representing the classification model. On the other hand, if the model generated in the training phase does not satisfy the generation conditions as a classification model, the generation unit 114 returns to the training phase, generates a new training dataset by increasing the number of datasets, and generates a classification model again using the new training dataset.

[0112] The generator 114 is mapped to a feature space corresponding to the feature quantities using a kernel method. The feature space before mapping is linear, but the kernel function maps it to a higher-dimensional feature space (the feature space after mapping) that has more dimensions. The generator 114 may perform mappings that extend dimensions using not only quantum kernels but also classical kernels.

[0113] (Regarding classification using a classification model) The classification unit 115 uses a classification model to identify whether each of the monitored objects T is normal or abnormal. The classification unit 115 inputs feature quantities of the sensor information to be classified into the classification model that has been verified in the verification phase and determined to satisfy the generation conditions. The feature quantities input into the classification model here are the feature quantities extracted by the feature quantity extraction unit 113 from the sensor information to be classified, which has been preprocessed by the preprocessing unit 112 after the sensor information to be classified has been acquired by the sensor information acquisition unit 111. The classification model classifies whether each of the monitored objects T10 and T2 is normal or abnormal based on the positional relationship between the input feature quantities and the separation boundary surface in the feature space, and outputs the classification result to the output unit 116.

[0114] (Regarding output) The output unit 116 outputs the classification result from the classification unit 115. The output unit 116 outputs the classification result to the display device 300. The display device 300 then displays the classification result. The output unit 116 may also output sensor information acquired by the sensor information acquisition unit 111 and the results of preprocessing performed by the preprocessing unit 112 (for example, the time-series waveform data shown in Figure 20) to the display device 300. The output unit 116 may also output features extracted by the feature extraction unit 113 to the display device 300. In this case, for example, the display device 300 displays the relationship between features and performance metrics (for example, a diagram like the one shown in Figure 25, which will be described later). The dataset storage unit 117 stores the dataset generated by the generation unit 114. The classification model storage unit 118 stores the classification model generated by the generation unit 114.

[0115] The storage unit of the information processing device 100 (including the dataset storage unit 117 and the classification model storage unit 118) is composed of storage media such as HDD (Hard Disk Drive), flash memory, EEPROM (Electrically Erasable Programmable Read Only Memory), RAM (Random Access read / write Memory), ROM (Read Only Memory), or a combination thereof. The storage unit of the information processing device 100 stores programs for executing various processes to realize the functions of the information processing device 100, and temporary data used when performing various processes.

[0116] The functional units of the information processing device 100 (including the sensor information acquisition unit 111, preprocessing unit 112, feature extraction unit 113, generation unit 114, classification unit 115, and output unit 116) are realized by causing the CPU (Central Processing Unit) and / or GPU (Graphics Processing Unit) provided as hardware in the information processing device 100 to execute programs stored in the memory unit of the information processing device 100.

[0117] (Regarding the processing flow performed by the information processing device 100) Here, the processing flow performed by the information processing device 100 will be explained using Figure 21. Figure 21 is a flowchart showing the processing flow performed by the information processing device 100 in the second embodiment.

[0118] First, the processing flow in the training phase will be explained. The preprocessing unit 112 of the information processing device 100 generates an AR model using the sensor information acquired by the sensor information acquisition unit 111 (step S1). Here, the sensor information used in the training phase is the sensor information when both the monitored targets T10 and T2 are in a normal state. The preprocessing unit 112 of the information processing device 100 extracts feature quantities (coefficients φ1, φ2, φ3, φ4, φ5 up to the 5th order) from the AR model generated in step S10 (step S2). The generation unit 114 of the information processing device 100 trains a One-Class SVM model with the training feature quantities (coefficients φ1, φ2, φ3, φ4, φ5 of the AR model) extracted in step S2, and generates a classification model MDL by training a model with an embedded kernel method (step S3). In the process of generating the classification model, the generation unit 114 generates a classification model MDL using a quantum kernel (step S4). Furthermore, the generation unit 114 generates a classification model MDL using a quantum kernel (step S5).

[0119] Next, the processing flow in the verification phase will be explained. The preprocessing unit 112 of the information processing device 100 generates an AR model using the sensor information acquired by the sensor information acquisition unit 111 (step S1). Here, the sensor information used in the verification phase consists of four sensor data points corresponding to the cases (normal / normal, abnormal / abnormal, abnormal / normal, normal / abnormal) for each of the two monitored targets T10 and T2. The generation unit 114 of the information processing device 100 verifies the classification model MDL generated in steps S3 to S5 (step S6). The generation unit 114 inputs the feature quantities (coefficients φ of the AR model) of the verification dataset into the classification model MDL. The generation unit 114 determines whether the classification result R output from the classification model MDL in response to the input of the verification dataset matches the labels of the verification dataset. If the result of classifying each of the verification datasets into the classification model MDL satisfies the generation conditions, the generation unit 114 terminates the generation of the classification model MDL.

[0120] Then, in the classification execution phase, similar to step S1, the preprocessing unit 112 generates an AR model using the sensor information acquired by the sensor information acquisition unit 111. The classification unit 115 inputs the feature quantities of the sensor information in the classification execution phase (coefficients φ of the AR model) into the classification model MDL. The classification unit 115 acquires the classification result R output from the classification model MDL as a classification result classified as normal or abnormal.

[0121] Here, the effects of the second embodiment will be explained using Figures 22 (Figures 22A, 22B), 23 (Figures 23A, 23B), and Figures 24 to 27. Figures 22 to 27 are diagrams illustrating the effects of the second embodiment.

[0122] Figure 22 shows the relationship between distance and sound pressure level in the measurement environment. Here, the distance corresponds to the x-coordinate of microphone M, which is positioned on the x-axis in Figure 18A. The sound pressure level is calculated based on the average value of five sound data sets collected for each condition.

[0123] Figure 22A shows the relationship between distance and sound pressure level for the sounds emitted by individual sound sources (conveyor CON, chain belt CHA, noise). The sound pressure level is calculated based on the average value of five sound data collections for each condition. Figure 22A shows the relationship between distance and sound pressure level for each of the following: CHA(0) representing a normal chain belt CHA, CHA(1) representing an abnormal chain belt CHA, CON(0) representing a normal conveyor CON, CON(1) representing an abnormal conveyor CON, WhiteNoise1 representing the first white noise, and WhiteNoise2 representing the second white noise. The first white noise is noise collected near the rubber belt machine Rub. The second white noise is noise collected in a corner of the measurement environment. In Figure 22A, for conveyor CON as the monitored object T10, the sound pressure level collected at a distance of 0m is shown to be 43.6 to 43.8 dB. Furthermore, for the conveyor CON, which is the monitored object T10, the sound pressure level decreased significantly as the distance increased, and the sound pressure levels collected at distances of 2m and 3m were shown to be 39.7–40.3dB. On the other hand, in Figure 22A, for the chain belt CHA, which is the monitored object T2, the sound pressure level collected at distance 0m was shown to be 40.4–42dB. Furthermore, for the chain belt CHA, which is the monitored object T2, the sound pressure level decreased significantly as the distance increased, and the sound pressure levels collected at distances of 2m and 3m were shown to be 36.4–36.8dB. When comparing the sound pressure levels for monitored objects T10 and T2, it was shown that, regardless of distance, the sound pressure level of the sound emitted from the conveyor CON was consistently higher than the sound pressure level of the sound emitted from the chain belt CHA. Furthermore, in the case of the chain belt CHA, there was a difference in sound pressure levels between normal and abnormal sounds at distances of 0m and 1m, making it possible to clearly distinguish between normal and abnormal sounds. At other locations (distances of 2m and 3m), no significant difference in sound pressure levels between normal and abnormal sounds was observed.Furthermore, at distances of 2m and 3m, the sound pressure levels of the chain belt CHA were shown to be close to those of the first and second white noises.

[0124] Figure 22B shows the relationship between distance and sound pressure level when two monitored devices T10 and T2 are operated in an environment with first white noise and second white noise. Figure 22B shows the relationship between distance and sound pressure level for each of the following: CON / CHA = 0 / 0, representing the operation of a normal conveyor CON and a normal chain belt CHA; CON / CHA = 1 / 0, representing the operation of an abnormal conveyor CON and a normal chain belt CHA; CON / CHA = 0 / 1, representing the operation of a normal conveyor CON and an abnormal chain belt CHA; CON / CHA = 1 / 1, representing the operation of an abnormal conveyor CON and an abnormal chain belt CHA; WhiteNoise1, representing the first white noise; and WhiteNoise2, representing the second white noise. In Figure 22B, it is shown that for each of CON / CHA = 0 / 0, 0 / 1, 1 / 0, and 1 / 1, the sound pressure level decreases as the distance increases. When both monitored objects T10 and T2 are operating, the combined sound pressure level (combined sound pressure) of the sound emitted from the conveyor CON and the chain belt CHA exceeds 80 dB. However, the sound pressure level measured in an environment where the conveyor CON and chain belt CHA were actually operating was approximately 46 dB. This indicates that the sound pressure level when both monitored objects T10 and T2 are operating is not the sum of the sound pressure levels of each monitored object T, but rather an increase of 4-5 dB from the sound pressure level of a single monitored object T. This phenomenon is thought to be due to interference caused by the overlapping of various sounds. It also indicates that the directional microphone M is effectively collecting sound in its direction of direction.

[0125] Figure 23 shows the relationship between performance metrics (accuracy and F1 score) and distance. Figure 23A shows the relationship between performance metrics (accuracy and F1 score) and distance for a classification model generated using a quantum kernel method. Figure 23B shows the relationship between performance metrics (accuracy and F1 score) and distance for a classification model generated using a conventional kernel method. Figures 23A and 23B show that the difference in performance metrics becomes particularly pronounced as the distance increases.

[0126] As shown in Figure 23A, the classification model generated using the quantum kernel method demonstrates very stable and high performance in both accuracy and F1 score, which are performance indicators.

[0127] Specifically, all of the tested microphones M10, M20, M30, and M4—that is, at distances of 0m, 1m, 2m, and 3m—maintained nearly optimal performance (for example, a performance index value of 0.9 or higher). When using sound data collected at a distance of 1m, a slight fluctuation was observed in the performance index, with an F1 score of approximately 0.92 and an accuracy of approximately 0.95. This slight degradation is thought to be due to acoustic interference. Particularly noteworthy is that, despite the sound pressure level decreasing with increasing distance and the proportion of noise to the sound corresponding to the normal / abnormal judgment (the sound emitted from each monitored object T) increasing, a performance index of approximately 1.0 was shown at a distance of 3m.

[0128] In contrast, Figure 23B shows that the performance index decreases significantly as the distance increases. Specifically, the performance index is approximately 1.0 at a distance of 0m, but it decreases sharply at a distance of 2m, and then increases at a distance of 3m, reaching approximately 0.60. At a distance of 0m, high accuracy and an F1 score (1.0) are shown for the sound collected from all channels CH1 to CH3. At a distance of 2m, the accuracy is 0.46 for channel CH1, 0.27 for channel CH2, and 0.41 for channel CH3, and the F1 score decreases sharply to 0.46 for channel CH1, 0.22 for channel CH2, and 0.42 for channel CH3. At a distance of 3m, the accuracy is approximately 0.72 to 0.73 for channels CH1 to CH3. At a distance of 3m, the F1 score also increases to approximately 0.69 for channels CH1 to CH3. This is thought to be due to sound interference caused by sound reflections and other factors in the measurement environment. However, the performance index at a distance of 0m is 1.0, and at this distance, a single sound source is dominant, indicating high performance. At a distance of 0m, it is thought that channel CH1 mainly collected sound generated from the chain belt CHA, channel CH3 mainly collected sound emitted from the conveyor CON, and channel CH2 collected sound emitted from both the chain belt CHA and the conveyor CON. As the distance increases, it is thought that each channel CH collects a complex mixture of sounds emitted from the chain belt CHA and the conveyor CON, as well as white noise.

[0129] The fact that classification models generated using quantum kernels consistently exhibit high performance despite increasing distance is thought to be due to the fundamental difference between classical and quantum feature spaces. In conventional classical kernels, the effectiveness of feature maps decreases as the signal-to-noise ratio decreases with increasing distance. Classical RBF kernels are effective in generating classification models with high classification accuracy under low-noise conditions (sound collection at a distance of 0m) because they can map separable data into the feature space. However, as the distance increases, environmental noise becomes dominant, and it becomes difficult to generate classification models with high classification accuracy because inseparable data is mapped into the feature space. On the other hand, quantum kernels can utilize the exponentially large Hilbert space available in quantum systems, making it possible to determine more complex and noise-resistant boundaries. Mapping data into a feature space using quantum kernels is thought to extract the original data pattern buried in noise by using coefficients of a identifiable AR model, even when the sound pressure level of the original data has attenuated and noise is dominant. Such resistance to noise and signal degradation due to distance can be a significant advantage in industrial environments where the placement of sensors such as microphones is often constrained by physical layout and safety considerations.

[0130] Figure 24 shows the results of an analysis of the performance differences in classification models generated using the quantum kernel method and the conventional classical kernel (Classical RBF). In Figure 24, the mean values ​​of the performance indicators for each of the three channels CH1 to CH3 are shown for each distance. A t-test revealed a statistically significant difference in the mean and standard deviation between the quantum kernel method and the conventional classical kernel (Classical RBF). For example, for accuracy, the t-test, p-value, and Cohen's d indicators are as follows: t(3) = 4.46, p = 0.021, Cohen's d = 1.49. Also, for the F1 score, the t-test, p-value, and Cohen's d indicators are as follows: t(3) = 4.81, p = 0.017, Cohen's d = 1.6. In the above, the values ​​significantly exceed the threshold (d = 0.8) established by Cohen (1988) as a d value that has a significant effect, statistically demonstrating that the quantum kernel method has a significant effect.

[0131] Figure 25 shows the confusion matrix in a classification model generated using the conventional classical kernel (Classical RBF) method. Figure 25 indicates that 30 to 34 false negatives (FPs) occurred for each channel CH. In actual manufacturing environments, such a high number of false positives is unacceptable.

[0132] Figure 26 shows the relationship between individual features and the performance metrics of the classification model. The horizontal axis of this figure represents the number of features, and the vertical axis represents the cumulative contribution rate. Figure 26 shows the relationship between five features—namely, the first-order coefficient φ1, the second-order coefficient φ2, the third-order coefficient φ3, the fourth-order coefficient φ4, and the fifth-order coefficient φ5 in the AR model—and the performance metrics of the classification model for both the quantum kernel method and the conventional classical kernel (Classical RBF). It is generally known that increasing the number of features and raising the cumulative contribution rate improves the performance of the model. In Figure 26, it is shown that the third feature (third-order coefficient φ3) and the fourth feature (fourth-order coefficient φ4) contribute significantly to the classification performance of the classification model. When a classification model is generated using a quantum kernel, it is shown that using five features allows the performance metrics (accuracy and F1 score) to reach approximately a constant level (1.0). This means that, in the case of a classification model using a quantum kernel, the cumulative contribution rate reaches 100% when using five features.

[0133] Figure 27 shows an example of mapping classification results from a classification model generated using a quantum kernel to a two-dimensional feature space using a third feature (cubic coefficient φ3) and a fourth feature (quaternary coefficient φ4). In Figure 27, the classification results for sensor information collected in measurement environments where CON / CHA = 0 / 0, 1 / 1, 0 / 1, and 1 / 0 are shown for each of the three channels CH (channels CH1 to CH3). The upper part of Figure 27 shows the classification results for channel CH3. When the conveyor CON is abnormal and the chain belt CHA is normal (CON / CHA = 1 / 0), it is shown that most of the classification results are mapped to the second quadrant in the two-dimensional feature space. When the conveyor CON is normal and the chain belt CHA is abnormal (CON / CHA = 0 / 1), it is shown that most of the classification results are mapped to the fourth quadrant in the two-dimensional feature space. When both the conveyor CON and the chain belt CHA are abnormal (CON / CHA = 1 / 1), a combined abnormality is shown, and the superposition of the above classification results is displayed, indicating that most of the classification results are mapped to the region from the second to the fourth quadrant.

[0134] As described above, the anomaly detection system 2 of the second embodiment comprises a sensor information acquisition unit 111, a preprocessing unit 112, a generation unit 114, and a classification unit 115. The sensor information acquisition unit 111 acquires sensor information. The sensor information is information detected by a number of sensors 200 that is less than the number of monitored objects T, each representing a state of a plurality of monitored objects T. The preprocessing unit 112 generates an AR model (autoregressive model) based on the time-series changes of the sensor information. The generation unit 114 generates a classification model. The classification model is a One-Class SVM (one-class support vector machine) that has learned the feature quantities of the AR model generated based on the sensor information corresponding to the normal state of each monitored object T. The learned model is a model that has been learned to classify whether each monitored object T is normal or abnormal based on the relationship between the feature quantities of the sensor information to be classified and the separation boundary, by learning the separation boundary between normal and abnormal for each monitored object T in the feature space formed by the feature quantities of the learned sensor information (sensor information corresponding to the normal state). The classification unit 115 classifies whether each monitored object is normal or abnormal by applying a classification model to the sensor information to be classified. As a result, the abnormality detection system 2 of the second embodiment can generate a classification model that classifies whether each monitored object T is normal or abnormal from sensor information measured by a number of sensors smaller than the number of monitored objects, and can use that classification model to classify whether each monitored object T is normal or abnormal. In other words, it is possible to detect whether an abnormality has occurred for each monitored object from sensor information measured by a number of sensors smaller than the number of monitored objects.

[0135] Furthermore, in the anomaly detection system 2 of the second embodiment, the generation unit 114 causes the One-Class SVM (One-Class Support Vector Machine) to learn the coefficients φi of the AR model. This allows the system to learn the features of the AR model.

[0136] Furthermore, the components of the anomaly detection system 2 of the second embodiment may be divided into a learning device and an anomaly detection device. The learning device is a device that executes the training phase and comprises at least a sensor information acquisition unit 111, a pre-processing unit 112, and a generation unit 114. The sensor information acquisition unit 111 acquires sensor information for learning. The pre-processing unit 112 generates an AR model based on the sensor information for learning. The generation unit 114 generates a classification model that has learned the features of the AR model. As a result, the learning device of the second embodiment can generate a classification model that classifies whether each of the monitored objects T is normal or abnormal from sensor information measured by a number of sensors smaller than the number of monitored objects. The anomaly detection device is a device that executes the classification execution phase and comprises at least a sensor information acquisition unit 111, a pre-processing unit 112, and a classification unit 115. The sensor information acquisition unit 111 acquires sensor information to be classified. The pre-processing unit 112 generates an AR model based on the sensor information to be classified. The classification unit 115 classifies whether each monitored object is normal or abnormal by applying a classification model to the feature quantities of the AR model generated based on the sensor information to be classified. As a result, in the abnormality detection device of the second embodiment, the classification model can be used to classify whether each monitored object T is normal or abnormal from sensor information measured by a number of sensors smaller than the number of monitored objects.

[0137] Application Examples of the Embodiments Here, we will describe application examples of the embodiments (the first and second embodiments). These application examples differ from the embodiments described above in that they perform quantum computation using a quantum computer provided by a cloud service.

[0138] The quantum computers provided here through cloud services are quantum computers that can be used on demand via the internet. Users can perform quantum computational processing without having to own an expensive quantum computer themselves.

[0139] Furthermore, quantum computation here refers to computational processing using a quantum kernel. More specifically, in the training phase of the first embodiment, the process by which the generation unit 14 uses a quantum kernel to generate a classification model based on feature quantities (sound time-series data) extracted from sound information as sensor information corresponds to quantum computation. Also, in the training phase of the second embodiment, the process by which the generation unit 114 uses a quantum kernel to generate a classification model based on feature quantities extracted from the AR model corresponds to quantum computation.

[0140] Figure 28 is a diagram illustrating application examples of the embodiments (first and second embodiments). In this diagram, the anomaly detection system 3 includes, for example, a factory F, a collaborative server NSV, a monitor MN, an information processing device 150, and a cloud CSV.

[0141] Factory F is a facility that has a manufacturing site W that is the target of anomaly detection. At Factory F, for example, sound information as a physical quantity generated from various devices operating at the manufacturing site W is collected by microphones. The sound information collected by the microphones is transmitted as acoustic data from Factory F to the linked server NSV. In addition, factory data, which is information about Factory F, is transmitted from Factory F to the linked server NSV. The factory data may include, for example, identification information of Factory F, identification information of the manufacturing site W collected by the microphones, and information indicating the date and time of sound collection.

[0142] CloudCSV is a quantum computer provided as a cloud service. CloudCSV performs computational processing using a quantum kernel in response to a request from the information processing device 150 and transmits the computation results to the information processing device 150.

[0143] The information processing device 150 is a computer corresponding to the information processing device 10 of the first embodiment or the information processing device 100 of the second embodiment. The information processing device 150 receives acoustic data collected at factory F via the interoperation server NSV. When the information processing device 150 functions as the information processing device 10 of the first embodiment, it transmits feature quantities (acoustic time-series data) extracted from the sound information to the cloud CSV and requests the cloud CSV to perform computational processing to generate a classification model based on the transmitted feature quantities (sound time-series data) using a quantum kernel. The cloud CSV generates a classification model in response to the request and transmits information indicating the generated classification model to the information processing device 150 as a computation result (calculation result). In other words, the generation unit 14 generates a classification model by causing the cloud CSV to perform computational processing using a quantum kernel. When the information processing device 150 functions as the information processing device 100 of the second embodiment, it transmits the feature quantities extracted from the AR model to the cloud CSV and requests the cloud CSV to perform computational processing to generate a classification model based on the transmitted feature quantities (features extracted from the AR model) using a quantum kernel. The cloud CSV generates a classification model in response to the request and transmits information indicating the generated classification model to the information processing device 150 as a computation result (calculation result). In other words, the generation unit 114 generates a classification model by causing the cloud CSV to perform computational processing using a quantum kernel. Thus, the information processing device 150 uses the external cloud CSV only when generating a classification model. For example, the information processing device 150 periodically (for example, every week) uses the cloud CSV for a few minutes to generate a classification model that has learned the separation boundary of sound information collected when the manufacturing equipment etc. is operating normally at the manufacturing site W. After the classification model is generated, the information processing device 150 performs normal / abnormal determination using the generated classification model internally. This makes it possible to perform quantum computing operations at a relatively low cost.

[0144] The NSV (Network Service Server) is a server that supports monitoring of the manufacturing site W. The NSV transmits acoustic data received from factory F to the information processing device 150. The NSV also receives judgment results from the information processing device 150 based on the acoustic data, that is, abnormal judgment results indicating whether the manufacturing equipment, etc., is operating normally or whether an abnormality has occurred at the manufacturing site W.

[0145] The NSV interoperation server monitors factory F based on the abnormality detection results received from the information processing device 150. For example, if the NSV interoperation server receives a result from the information processing device 150 indicating that the manufacturing equipment is operating normally, it allows the manufacturing site W to continue operating. On the other hand, if the NSV interoperation server receives a result from the information processing device 150 indicating that an abnormality has occurred in the manufacturing equipment, it issues a notification prompting the factory to take action to resolve the abnormal condition, such as stopping the operation of the manufacturing site W.

[0146] The NSV interoperation server outputs the abnormality detection result received from the information processing device 150 to the monitor MN. The NSV interoperation server may also output the acoustic data and factory data received from factory F to the monitor MN.

[0147] The monitor MN is a computer corresponding to the display device 30 of the first or second embodiment. The monitor MN displays the abnormality detection result received from the cooperating server NSV. The monitor MN may also display the received acoustic data and factory data received from the cooperating server NSV. The images displayed on the monitor MN, i.e., the abnormality detection result, acoustic data, and factory data, are visually inspected by the user U, such as a supervisor or operator.

[0148] As described above, the anomaly detection system 3 according to the modified embodiments (first and second embodiments) further includes a cloud CSV that performs computational processing using a quantum kernel in response to requests from users (on demand). The information processing device 150 requests the cloud CSV to generate a classification model by transmitting sensor information or feature quantities of sensor information, and receives information on the classification model generated by the cloud CSV in response to this request from the cloud CSV. The information processing device 150 performs anomaly determination using the classification model received from the cloud CSV. As a result, the anomaly detection system 3 according to the modified embodiments (first and second embodiments) can perform computational processing related to quantum computing at a relatively low cost.

[0149] The anomaly detection systems 1 to 3 and information processing devices 10, 100, and 150 in the above-described embodiments (first and second embodiments) and their modified versions may be implemented in whole or in part using a computer. In that case, the functions may be implemented by recording a program for implementing these functions on a computer-readable recording medium, loading the program recorded on this recording medium into a computer system, and executing it. Here, "computer system" includes hardware such as an OS and peripheral devices. Furthermore, "computer-readable recording medium" refers to portable media such as flexible disks, magneto-optical disks, ROMs, CD-ROMs, and storage devices such as hard disks built into a computer system. Moreover, "computer-readable recording medium" may also include those that dynamically hold programs for a short period of time, such as communication lines used when transmitting programs via networks such as the Internet or communication lines such as telephone lines, and those that hold programs for a certain period of time, such as volatile memory inside a computer system that acts as a server or client in such cases. Furthermore, the above-mentioned program may be for implementing a part of the functions described above, or it may be a program that can implement the above-mentioned functions in combination with a program already recorded in the computer system, or it may be implemented using a programmable logic device such as an FPGA.

[0150] While embodiments (first and second embodiments) and their modifications have been described in detail above with reference to the drawings, the specific configuration is not limited to these embodiments (first and second embodiments) and their modifications, and includes designs and the like that do not depart from the spirit of the invention.

[0151] 1, 2, 3... Anomaly detection system 10, 100, 150... Information processing device 11, 111... Sensor information acquisition unit 12, 112... Preprocessing unit 13, 113... Feature extraction unit 14, 114... Generation unit 15, 115... Classification unit 16, 116... Output unit 17, 117... Dataset storage unit 18, 118... Classification model storage unit 20, 200... Sensor

Claims

1. An anomaly detection system comprising: a sensor information acquisition unit that acquires sensor information by detecting the state of each of multiple monitored targets with a number of sensors less than the number of monitored targets; a generation unit that generates a classification model which is a One-Class SVM (One-Class Support Vector Machine) that has learned the sensor information corresponding to the normal state of each monitored target, and which is trained to classify whether each monitored target is normal or abnormal based on the relationship between the feature quantities of the sensor information to be classified and the separation boundary in the feature space formed by the features of the learned sensor information; and a classification unit that classifies whether each monitored target is normal or abnormal by applying the classification model to the sensor information of the target to be classified acquired by the sensor information acquisition unit.

2. The anomaly detection system according to claim 1, wherein the generation unit generates sensor information corresponding to the normal state of each of the plurality of monitored targets by synthesizing the individual detection information obtained when each of the monitored targets is detected as being in a normal state by a sensor, and causes a One-Class SVM (One-Class Support Vector Machine) to learn the generated sensor information.

3. The anomaly detection system according to claim 1, wherein the generation unit learns a separation boundary for classifying features obtained by mapping features extracted from the sensor information to a feature space with a higher dimension than the number of dimensions of the features using a kernel method.

4. The anomaly detection system according to claim 3, wherein the generation unit learns a separation boundary for classifying the mapped feature quantities using a quantum kernel created by a quantum kernel circuit.

5. The anomaly detection system according to claim 1, wherein the sensor information is information detected by a non-contact sensor.

6. The anomaly detection system according to claim 5, wherein the sensor information is sound information obtained by collecting sounds emitted by each of the plurality of monitored objects with a microphone and synthesizing the collected sounds, and further comprises: a preprocessing unit that extracts sound components for each frequency component in the sound information by filtering the sound information; and a feature extraction unit that uses the principal components extracted by principal component analysis of the sound components for each frequency component in the sound information generated by the preprocessing unit as feature quantities of the sound information, and the generation unit learns a separation boundary for classifying feature quantities mapped to the feature quantities of the sound information using a quantum kernel created with a quantum kernel circuit.

7. A learning device comprising: a sensor information acquisition unit that acquires sensor information by detecting the state of each of multiple monitored targets with a number of sensors less than the number of monitored targets; and a generation unit that generates a classification model which is a One-Class SVM (One-Class Support Vector Machine) that has learned the sensor information corresponding to the normal state of each monitored target as an unsupervised machine learning, and which learns the separation boundary between normal and abnormal for each monitored target in the feature space formed by the features of the learned sensor information, thereby classifying whether each monitored target is normal or abnormal based on the relationship between the features of the sensor information to be classified and the separation boundary.

8. An anomaly detection device comprising: a sensor information acquisition unit that acquires sensor information by detecting the state of each of multiple monitored targets with a number of sensors less than the number of monitored targets; and a classification unit that classifies whether each monitored target is normal or abnormal by applying a classification model to the sensor information acquired by the sensor information acquisition unit, wherein the classification model is a One-Class SVM (One-Class Support Vector Machine) that has learned the sensor information corresponding to the normal state of each monitored target, and is a model that has been learned to classify whether each monitored target is normal or abnormal based on the relationship between the feature quantities of the sensor information of the target to be classified and the separation boundary, by learning the separation boundary between normal and abnormal for each monitored target in the feature space formed by the features of the learned sensor information.

9. A learning method performed by a computer used in an anomaly detection system, wherein a sensor information acquisition unit acquires sensor information by detecting the state of each of a plurality of monitored objects with a number of sensors less than the number of monitored objects, and a generation unit is a One-Class SVM (One-Class Support Vector Machine) that has learned the sensor information corresponding to the normal state of each monitored object as unsupervised machine learning, and generates a classification model that has learned to classify whether each monitored object is normal or abnormal based on the relationship between the features of the sensor information to be classified and the separation boundary of each monitored object in the feature space formed by the features of the learned sensor information.

10. An anomaly detection method performed by a computer used in an anomaly detection system, wherein a sensor information acquisition unit acquires sensor information by detecting the state of each of a plurality of monitored objects with a number of sensors less than the number of monitored objects, and a classification unit classifies each monitored object as normal or abnormal by applying a classification model to the sensor information acquired by the sensor information acquisition unit, and the classification model is a One-Class SVM (One-Class Support Vector Machine) that has learned the sensor information corresponding to the normal state of each monitored object, and is a model that has been learned to classify each monitored object as normal or abnormal based on the relationship between the feature quantities of the sensor information of the object to be classified and the separation boundary, by learning the separation boundary between normal and abnormal for each monitored object in the feature space formed by the features of the learned sensor information.

11. A program for an anomaly detection system that uses a computer to acquire sensor information by detecting the state of each of multiple monitored targets with fewer sensors than the number of monitored targets, and a One-Class SVM (One-Class Support Vector Machine) that has learned the sensor information corresponding to the normal state of each monitored target, and generates a classification model that is trained to classify whether each monitored target is normal or abnormal based on the relationship between the features of the sensor information to be classified and the separation boundary of each monitored target in the feature space formed by the features of the learned sensor information, and then applies the classification model to the acquired sensor information to be classified to classify whether each monitored target is normal or abnormal.

12. A One-Class Support Vector Machine (SVM) used in an anomaly detection system, which acquires sensor information from a number of sensors fewer than the number of monitored objects, detecting the state of each of the multiple monitored objects, and learns the sensor information corresponding to the normal state of each monitored object as an unsupervised machine learning. The program generates a classification model that learns the separation boundary between normal and abnormal states for each monitored object in the feature space formed by the features of the learned sensor information, and then classifies whether each monitored object is normal or abnormal based on the relationship between the features of the sensor information to be classified and the separation boundary.

13. A program for a computer used in an anomaly detection system, which acquires sensor information by detecting the state of each of multiple monitored objects with a number of sensors less than the number of monitored objects, and classifies each monitored object as normal or abnormal by applying a classification model to the acquired sensor information, wherein the classification model is a One-Class SVM (One-Class Support Vector Machine) that has learned the sensor information corresponding to the normal state of each monitored object, and is a model that has been learned to classify each monitored object as normal or abnormal based on the relationship between the features of the sensor information of the object to be classified and the separation boundary, by learning the separation boundary between normal and abnormal for each monitored object in the feature space formed by the features of the learned sensor information.

14. An anomaly detection system comprising: a sensor information acquisition unit that acquires sensor information by detecting the state of each of multiple monitored targets with a number of sensors less than the number of monitored targets; a preprocessing unit that generates an autoregressive model based on the time-series changes of the sensor information acquired by the sensor information acquisition unit; a generation unit that generates a classification model which is a One-Class SVM (One-Class Support Vector Machine) that has learned the feature quantities of the autoregressive model generated based on the sensor information corresponding to the normal state of each monitored target, and which is trained to classify whether each monitored target is normal or abnormal based on the relationship between the feature quantities of the sensor information of the target to be classified and the separation boundary in the feature space formed by the learned feature quantities of the sensor information; and a classification unit that classifies whether each monitored target is normal or abnormal by applying the classification model to the feature quantities of the autoregressive model generated based on the sensor information of the target to be classified acquired by the sensor information acquisition unit.

15. The anomaly detection system according to claim 14, wherein the generation unit causes a One-Class SVM (One-Class Support Vector Machine) to learn the coefficients of the autoregressive model.

16. The anomaly detection system according to claim 14, wherein the sensor information is information detected by a non-contact sensor.

17. The anomaly detection system according to claim 16, wherein the sensor information is sound information obtained by collecting sounds emitted by each of the plurality of monitored objects using a microphone.

18. A learning device comprising: a sensor information acquisition unit that acquires sensor information by detecting the state of each of multiple monitored targets with a number of sensors less than the number of monitored targets; a preprocessing unit that generates an autoregressive model based on the time-series changes of the sensor information acquired by the sensor information acquisition unit; and a generation unit that generates a classification model which is a One-Class SVM (One-Class Support Vector Machine) that has learned the features of the autoregressive model generated based on the sensor information corresponding to the normal state of each monitored target as an unsupervised machine learning method, and which learns the separation boundary between normal and abnormal for each monitored target in the feature space formed by the learned sensor information features, thereby classifying whether each monitored target is normal or abnormal based on the relationship between the sensor information features of the target to be classified and the separation boundary.

19. An anomaly detection device comprising: a sensor information acquisition unit that acquires sensor information by detecting the state of each of multiple monitored targets with a number of sensors less than the number of monitored targets; a preprocessing unit that generates an autoregressive model based on the time-series changes of the sensor information acquired by the sensor information acquisition unit; and a classification unit that classifies whether each monitored target is normal or abnormal by applying a classification model to the features of the autoregressive model generated based on the sensor information acquired by the sensor information acquisition unit, wherein the classification model is a One-Class SVM (One-Class Support Vector Machine) that has learned the features of the autoregressive model generated based on the sensor information corresponding to the normal state of each monitored target, and is a model that has been learned to classify whether each monitored target is normal or abnormal based on the relationship between the features of the sensor information of the target to be classified and the separation boundary, by learning the separation boundary between normal and abnormal for each monitored target in the feature space formed by the learned sensor information features.

20. A learning method performed by a computer used in an anomaly detection system, comprising: a sensor information acquisition unit acquiring sensor information by detecting the state of each of a plurality of monitored targets with a number of sensors less than the number of monitored targets; a preprocessing unit generating an autoregressive model based on the time-series changes of the sensor information acquired by the sensor information acquisition unit; and a generation unit generating a One-Class SVM (One-Class Support Vector Machine) that has learned the features of the autoregressive model generated based on the sensor information corresponding to the normal state of each monitored target as unsupervised machine learning, and generating a classification model that has learned to classify whether each monitored target is normal or abnormal based on the relationship between the features of the sensor information to be classified and the separation boundary of each monitored target in the feature space formed by the learned sensor information features, the learning method.

21. An anomaly detection method performed by a computer used in an anomaly detection system, comprising: a sensor information acquisition unit acquiring sensor information by detecting the state of each of a plurality of monitored targets with a number of sensors less than the number of monitored targets; a preprocessing unit generating an autoregressive model based on the time-series changes of the sensor information acquired by the sensor information acquisition unit; a classification unit classifying each monitored target as normal or abnormal by applying a classification model to the features of the autoregressive model generated based on the sensor information acquired by the sensor information acquisition unit; and a classification model being a One-Class SVM (One-Class Support Vector Machine) that has learned the features of the autoregressive model generated based on the sensor information corresponding to the normal state of each monitored target, and a model that has been learned to classify each monitored target as normal or abnormal based on the relationship between the features of the sensor information of the target to be classified and the separation boundary in the feature space formed by the learned sensor information features.

22. A One-Class Support Vector Machine (SVM) used in an anomaly detection system, which involves a computer that acquires sensor information by detecting the state of each of multiple monitored targets using fewer sensors than the number of monitored targets, generates an autoregressive model based on the time-series changes of the acquired sensor information, and learns the features of the autoregressive model generated based on the sensor information corresponding to the normal state of each monitored target, and generates a classification model trained to classify whether each monitored target is normal or abnormal based on the relationship between the features of the sensor information of the target to be classified and the separation boundary in the feature space formed by the learned sensor information features, and then applies the classification model to the features of the autoregressive model generated based on the acquired sensor information of the target to be classified, thereby classifying whether each monitored target is normal or abnormal.

23. A One-Class Support Vector Machine (SVM) used in an anomaly detection system, which involves a computer that acquires sensor information by detecting the state of each of multiple monitored targets using fewer sensors than the number of monitored targets, generating an autoregressive model based on the time-series changes of the acquired sensor information, and learning the features of the autoregressive model generated based on the sensor information corresponding to the normal state of each monitored target as an unsupervised machine learning method, and generating a classification model that learns the separation boundary between normal and abnormal states of each monitored target in the feature space formed by the learned sensor information features, thereby classifying whether each monitored target is normal or abnormal based on the relationship between the sensor information features of the target to be classified and the separation boundary.

24. A program for a computer used in an anomaly detection system, which acquires sensor information by detecting the state of each of multiple monitored objects with a number of sensors less than the number of monitored objects; generates an autoregressive model based on the time-series changes of the acquired sensor information; and classifies each monitored object as normal or abnormal by applying a classification model to the features of the autoregressive model generated based on the acquired sensor information; the classification model is a One-Class SVM (One-Class Support Vector Machine) that has learned the features of the autoregressive model generated based on the sensor information corresponding to the normal state of each monitored object, and is a model that has been learned to classify each monitored object as normal or abnormal based on the relationship between the features of the sensor information of the object to be classified and the separation boundary, by learning the separation boundary between normal and abnormal for each monitored object in the feature space formed by the learned sensor information features.

25. An anomaly detection system according to claim 1 or claim 14, further comprising a cloud server that performs computational processing using a quantum kernel on demand in response to a user request, wherein the generation unit generates the classification model by causing the cloud server to perform computational processing using a quantum kernel.

Images