Data composition discrimination device, method and program

Abstract

To provide a data composition discrimination device, method and program capable of highly accurately discriminating a data composition in which data components from a plurality of data sources are mixed.SOLUTION: A data composition discrimination device according to an embodiment includes a first acquisition unit, a second acquisition unit, a calculation unit and a discrimination unit. The first acquisition unit acquires mixed data of a discrimination object in which data components from a plurality of data sources are mixed. The second acquisition unit acquires a model including a plurality of layers from an input layer to an output layer. The calculation unit applies the mixed data to the model, and calculates characteristics of the mixed data for each layer of a portion of or all of the plurality of layers. The discrimination unit discriminates a composition of data sources of data components constituting the mixed data on the basis of the calculated characteristics.SELECTED DRAWING: Figure 5

Description

[Technical Field]

[0001] An embodiment of the present invention relates to a data composition determination device, method, and program. [Background technology]

[0002] Separating input data that contains a mixture of multiple data components into individual data components is an important problem that is expected to be applied in a wide range of situations, such as segmentation, which divides an image into various objects or backgrounds that appear in the image, or signal separation, which extracts anomalous signals buried in noise. In recent years, with the development of statistical machine learning models, especially deep learning models, more and more accurate methods are being proposed every day.

[0003] However, to fully utilize the performance of a machine learning model, the data used to train the model and the actual data to be processed must belong to the same domain. Especially when operating a model over a long period of time, environmental changes over time can gradually change the characteristics of the data, potentially exceeding the range in which the model functions properly. In the case of deep learning models, coupled with their black-box nature, it is often difficult to identify the factors that cause a deterioration in model performance. In particular, when processing input data that contains a mixture of data components from multiple data sources, there are multiple objects that can change over time, making it even more important to address deterioration due to changing data trends. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent Publication No. 2021-117800 Summary of the Invention [Problem to be solved by the invention]

[0005] The problem to be solved by the present invention is to provide a data composition determination device, method, and program that can determine with high accuracy the composition of data in which data components from multiple data sources are mixed. [Means for solving the problem]

[0006] A data composition determination device according to an embodiment includes a first acquisition unit, a second acquisition unit, a calculation unit, and a determination unit. The first acquisition unit acquires mixed data to be determined, which is a mixture of data components from multiple data sources. The second acquisition unit acquires a model including multiple layers from an input layer to an output layer. The calculation unit applies the mixed data to the model and calculates features of the mixed data for some or all of the multiple layers. The determination unit determines the data source composition of the data components that make up the mixed data based on the calculated features. [Brief explanation of the drawings]

[0007] [Figure 1] FIG. 1 is a diagram illustrating an example of a hardware configuration of a data composition determination device. [Figure 2] Diagram showing the structure of mixed data [Figure 3] Diagram showing an example of mixed data [Figure 4] A table showing the classification of data composition when N=1 [Figure 5] FIG. 1 shows the processing steps of a data composition determination method. [Figure 6] A diagram showing the relationship between parts in the data composition determination method shown in Figure 5. [Figure 7] Diagram showing the discrimination process for six categories of data composition [Figure 8] FIG. 10 is a diagram showing an example of a display screen of the determination result of data composition. [Figure 9] FIG. 10 is a diagram showing an evaluation of the discrimination result of data composition according to the present embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0008] Hereinafter, a data composition determination device, method, and program according to this embodiment will be described with reference to the drawings.

[0009] Fig. 1 is a diagram showing an example of the hardware configuration of a data composition determination device 100. As shown in Fig. 1, the data composition determination device 100 is a computer having a processor 1, a storage device 2, an input device 3, a display device 4, and a communication device 5. Transmission and reception of data and various signals between the processor 1, the storage device 2, the input device 3, the display device 4, and the communication device 5 is performed via a bus.

[0010] The processor 1 is an integrated circuit that controls the overall operation of the data composition determination device 100. For example, the processor 1 has a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), a DSP (Digital Signal Processor), and / or an FPU (Floating-Point Unit). The processor 1 may also have an internal memory and an I / O interface. The processor 1 executes various processes by interpreting and calculating programs stored in advance in a storage device 2 or the like. The processor 1 may also be realized in part or in whole by hardware such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

[0011] The storage device 2 is a volatile memory and / or a non-volatile memory that stores various data. For example, the storage device 2 stores data and setting values ​​used when the processor 1 executes various processes, data generated by various processes in the processor 1, etc. The storage device 2 is configured with a ROM (Read Only Memory), a RAM (Random Access Memory), an HDD (Hard Disk Drive), an SSD (Solid State Drive), an integrated circuit storage device, etc. The storage device 2 may also include a non-transitory computer-readable storage medium that stores a program executed by the processor 1.

[0012] The input device 3 accepts various operation inputs from an operator. Examples of the input device 3 that can be used include a keyboard, a mouse, various switches, a touchpad, and a touch panel display. An electrical signal corresponding to the accepted operation input (hereinafter referred to as an operation signal) is supplied to the processor 1.

[0013] The display device 4 displays various data under the control of the processor 1. A CRT (Cathode-Ray Tube) display, a liquid crystal display, an organic EL (Electro Luminescence) display, an LED (Light-Emitting Diode) display, a plasma display, or any other display may be used as appropriate as the display device 4. The display device 4 may also be a projector.

[0014] The communication device 5 includes a communication interface such as a network interface card (NIC) for performing data communication with various devices connected to the data composition determination device 100 via a network. Note that an operation signal may be supplied from a computer connected via the communication device 5 or an input device provided in the computer, and various data may be displayed on a display device or the like provided in the computer connected via the communication device 5. However, for the sake of simplicity of the following explanation, unless otherwise specified, it is assumed that the source of the operation signal is the input device 3 and the display destination of the various data is the display device 4. The input device 3 can be replaced by a computer connected via the communication device 5 or an input device provided in the computer, and the display device 4 can be replaced by a display device or the like provided in the computer connected via the communication device 5.

[0015] The data composition determination device 100 does not need to include all of the processor 1, storage device 2, input device 3, display device 4, and communication device 5. Depending on the design, some of the storage device 2, input device 3, display device 4, and communication device 5 may not be provided. The data composition determination device 100 may be provided with any additional hardware device useful for executing the processing of this embodiment. The data composition determination device 100 does not need to be physically composed of a single computer, but may be composed of a computer system having multiple computers connected to each other for communication via wires or a network line. The allocation of the series of processing according to this embodiment to the multiple processors 1 implemented in each of the multiple computers can be arbitrarily set. All processors 1 may execute all processing in parallel, or specific processing may be assigned to one or some of the processors 1, and the series of processing according to this embodiment may be executed by the entire computer system.

[0016] As shown in FIG. 1, the processor 1 functions as a mixed data acquisition unit 11, a model acquisition unit 12, a feature calculation unit 13, a composition determination unit 14, and a display control unit 15 by executing the above program.

[0017] The mixed data acquisition unit 11 acquires various mixed data. "Mixed data" refers to data in which data components from multiple data sources are mixed. "Data" includes image data, which is a series of pixel value data; audio data, which is a series of data on the amplitude of an audio signal; text data, which is a series of data on letters, numbers, symbols, etc.; and sensor data, which is a series of sensor output values. As sensors, voltage sensors, current sensors, power sensors, vibration sensors, magnetic sensors, weight sensors, speed sensors, acceleration sensors, distance sensors, and other sensors that measure any physical quantity can be applied. "Data source" refers to the source of the various data components contained in the data.

[0018] FIG. 2 is a diagram illustrating the structure of mixed data. As shown in FIG. 2, each of the multiple data sources included in the mixed data belongs to either a foreground data source group or a background data source group. The foreground data source group has N (N is a natural number greater than or equal to 0) foreground data sources. In the foreground data source group, each of the N foreground data sources is individually determined to determine whether or not it contains a data component of that foreground data source and whether or not the trend of the data component of that foreground data source has changed. As an example, an abnormal location captured in image data or an abnormal component included in time-series data corresponds to a data component of the foreground data source. An abnormal component occurs non-stationarily during the measurement period of the mixed data. A situation in which it is desired to individually determine the data components of all data sources that may be included in the mixed data corresponds to a situation in which all data sources belong to the foreground data source group.

[0019] As shown in Figure 2, the background data source group has M (M is a natural number greater than or equal to 1) background data sources. Data components of the background data sources are constantly generated during the measurement period of the mixed data. In the background data source group, each of the M background data sources is not individually determined to be included in the data components of that background data source, but it is determined whether the trend of any of the data components of the M background data sources has changed. For example, background components that inevitably enter image data or noise components that are always included in time-series data do not need to be individually determined; it is sufficient to recognize whether any change has occurred overall. Assuming such a situation, the background data source group is a collection of data sources in which it is known in advance that at least one of the multiple data sources will always be included in the mixed data, and it is not necessary to determine whether that data source exists, but it is necessary to determine whether any change has occurred in any of the data sources. Since it is not necessary to understand the breakdown of the background data source group, the entire background data source group can be considered to constitute a single data source. For example, this would include situations where the sky and forest in the background of image data are not distinguished and all are processed together as background, or where multiple types of noise components contained in time-series data are not distinguished and are processed together as a single type of noise component. Considering the above, a situation where all data sources that can be included in mixed data belong to the background data source group is equivalent to detecting a trend change for a single type of data input.

[0020] FIG. 3 is a diagram showing an example of mixed data. In FIG. 3, the mixed data is sensor data from a voltage sensor for detecting discharge, and is data that represents changes in voltage value (Voltage) for each position (Position). The mixed data includes a mixture of partial discharge (PD) components and noise components. That is, the mixed data shown in FIG. 3 is an example of N=1 and M=1. The partial discharge components are classified into a foreground data source group and are data components whose foreground data source is a partial discharge source. The noise components are classified into a background data source group and are data components whose background data source is a noise source. The noise components occur steadily during the measurement period of the mixed data. The partial discharge components are small, non-stationary data components buried in strong noise components and are more localized than the noise components. The partial discharge components occur steadily during the measurement period of the mixed data. The mixed data may include data components from other foreground data sources and background data sources. Which data source is the foreground data source and which is the background data source can be arbitrarily set according to interest.

[0021] As another example, the mixed data may be image data of a cityscape. In this case, the mixed data may include, for example, data components such as people, buildings, roads, and sky, where the person data component is foreground data with the person as the foreground data source, and the building data component is foreground data with the building as the foreground data source. Furthermore, the road data component is background data with the road as the background data source, and the sky data component is background data with the sky as the background data source.

[0022] As another example, the mixed data may be text data relating to multiple topics. In this case, the mixed data may include sentences describing multiple topics, with various word data components being foreground data with the words as foreground data sources, preposition data components being background data with the prepositions as background data sources, and context data components being background data with the contexts as background data sources.

[0023] Specifically, the mixed data acquisition unit 11 acquires mixed data to be distinguished, which is a mixture of data components from multiple data sources. The mixed data to be distinguished means data whose composition of data components is unknown. In addition, for each of the multiple foreground data sources, the mixed data acquisition unit 11 acquires first other mixed data in which a data source other than the foreground data source is dominant, and second other mixed data in which the foreground data source is dominant. The mixed data acquisition unit 11 generates each of the first other mixed data and the second other mixed data based on known mixed data.

[0024] The model acquisition unit 12 acquires a model including multiple layers from an input layer to an output layer. The model is assumed to be a deep learning model, which is a multi-layered machine learning model trained by machine learning to perform specific information processing. The type of information processing is not particularly limited as long as it extracts features from input data, and examples that can be used include noise reduction (denoising), defect extraction, segmentation, sound source separation, and related sentence search.

[0025] The feature calculation unit 13 applies the mixed data to be discriminated acquired by the mixed data acquisition unit 11 to a model in the model acquisition unit 12, and calculates the features of the mixed data for each of some or all of the multiple layers. Similarly, the feature calculation unit 13 also applies each of the first other mixed data and second other mixed data acquired by the mixed data acquisition unit 11 to a model, and calculates the features of the mixed data for each of some or all of the multiple layers.

[0026] The composition determination unit 14 determines the composition of data sources (hereinafter referred to as "data composition") of data components constituting the mixed data to be determined based on the features calculated by the feature calculation unit 13. Specifically, the composition determination unit 14 includes a signal source determination module, a first change detection module, and a second change detection module. The signal source determination module determines, for each of one or more foreground data sources, whether the foreground data source is included in the mixed data based on features from a first designated layer among the multiple layers. The first change detection module determines, for each of one or more foreground data sources, whether the trend of the data components of the foreground data source included in the mixed data has changed based on features from a second designated layer among the multiple layers. The second change detection module determines whether the trend of the data components of the foreground data source included in the mixed data has changed based on features from a third designated layer among the multiple layers, and combines the determination results based on the features from the third designated layer for one or more foreground data sources to determine whether the trend of the data components of the background data sources included in the mixed data has changed.

[0027] The data composition is classified based on a combination of whether or not the mixed data contains a foreground data source (discrimination result D1), whether or not the trend of the data components of the foreground data source has changed (discrimination result D2), and whether or not the trend of the data components of the background data source group contained in the mixed data has changed (discrimination result D3).

[0028] Fig. 4 is a diagram showing in tabular form the classification of data compositions when N = 1. As shown in Fig. 4, the data compositions are classified into six data compositions based on the combination of the discrimination results D1, D2, and D3. Specifically, the mixed data is classified into the following: a first composition (KB / N) in which the mixed data contains only data components of known background data sources (no data components of foreground data sources); a second composition (KB / KF) in which the mixed data contains data components of known background data sources and data components of known foreground data sources; a third composition (KB / UF) in which the mixed data contains data components of known background data sources and data components of unknown foreground data sources; a fourth composition (UB / N) in which the mixed data contains only data components of unknown background data sources (no data components of foreground data sources); a fifth composition (UB / KF) in which the mixed data contains data components of unknown background data sources and data components of known foreground data sources; and a sixth composition (UB / UF) in which the mixed data contains data components of unknown background data sources and data components of unknown foreground data sources.

[0029] The first composition (KB / N) corresponds to a case where the discrimination result D1 indicates that there is no foreground data source, and the discrimination result D3 indicates that the background data source group is known. Regarding discrimination result D3, a change in the trend of the data components of the background data source group means that the background data source group is an unknown background that is qualitatively different from known data sources, while a lack of change in the trend of the data components of the background data source group means that the background data source group is of the same quality as known data sources. The second composition (KB / KF) corresponds to a case where the discrimination result D1 indicates that there is a foreground data source, the discrimination result D2 indicates that the foreground data source is known, and the discrimination result D3 indicates that the background data source group is known. Regarding discrimination result D2, a change in the trend of the data components of the foreground data source means that the foreground data source is unknown, while a lack of change in the trend of the data components of the foreground data source means that the foreground data source is known.

[0030] Similarly, the third composition (KB / UF) corresponds to the case where it is determined that the discrimination result D1 shows a foreground data source, the discrimination result D2 shows an unknown foreground data source, and the discrimination result D3 shows a known background data source group. The fourth composition (UB / N) corresponds to the case where it is determined that the discrimination result D1 shows no foreground data source, and the discrimination result D3 shows an unknown background data source group. The fifth composition (UB / KF) corresponds to the case where it is determined that the discrimination result D1 shows a foreground data source, the discrimination result D2 shows a known foreground data source, and the discrimination result D3 shows an unknown background data source group. The sixth composition (UB / UF) corresponds to the case where it is determined that the discrimination result D1 shows a foreground data source, the discrimination result D2 shows an unknown foreground data source, and the discrimination result D3 shows an unknown background data source group. Data composition in a general N is a 2×3 matrix based on a combination of 2N+1 discrimination results. N It is classified as a street.

[0031] The display control unit 15 displays various information. As one example, the display control unit 15 displays the data composition determined by the composition determination unit 14 on the display device 4. As another example, the display control unit 15 may display the mixed data to be determined, the first other mixed data, and / or the second other mixed data acquired by the mixed data acquisition unit 11 on the display device 4.

[0032] Next, an example of the operation of the data composition determination device 100 according to this embodiment will be described.

[0033] In the following embodiments, the mixed data is assumed to be composed of a foreground data source group and a background data source group. Consider cases where there are two or more data sources in the foreground data source group, or where there is one or more data sources in each of the foreground data source group and the background data source group. However, even in cases other than these, it is possible to determine the composition of the mixed data using the following method.

[0034] FIG. 5 is a diagram showing the processing procedure of the data composition determination method by the processor 1. FIG. 6 is a diagram showing the relationship between the units in the data composition determination method shown in FIG. 5. As shown in FIGS. 5 and 6, the mixed data acquisition unit 11 acquires mixed data X to be determined (step SA1). The mixed data X includes data components from data sources belonging to the foreground data source group and data components from data sources belonging to the background data source group. Multiple mixed data X may be acquired.

[0035] After step SA1 is performed, the mixed data acquisition unit 11 acquires mixed data Yi in which data source i is the main data source (step SA2). Here, data source i refers to the foreground data source to be identified. The foreground data source to be identified is selected according to an arbitrary rule from one or more predetermined foreground data sources that may be included in the foreground data source group. The following steps SA2 to SA9 are repeated for each data source i.

[0036] The acquisition of the mixed data Yi will now be described in detail. The storage device 2 stores a large number of known data items related to a large number of data sources. Assume that this known data was used to train the individual model Mi for data source i. The mixed data acquisition unit 11 generates the mixed data Yi by arbitrarily combining the known data items stored in the storage device 2. As an example, a combination of data sources designated so that data source i is dominant is determined by a user or the like. The designated combination is composed of data source i and data sources other than data source i, and the distribution of data source i and the other data sources may be arbitrarily determined so that data source i is the most dominant component and the background data source is the second most dominant component. Next, the mixed data acquisition unit 11 reads known data corresponding to the combination from the storage device 2 and combines the read known data. This generates mixed data Yi in which data source i is dominant. The mixed data Yi corresponds to the data before the trend change. The mixed data Yi corresponds to a second, other mixed data. Multiple pieces of mixed data Yi may be generated as long as the above conditions are met.

[0037] When step SA2 is performed, the mixed data acquisition unit 11 acquires mixed data Zi in which the background data source group, in particular, is the main component other than the data source i (step SA3). Like the mixed data Yi, the mixed data Zi can be generated by arbitrarily combining known data stored in the storage device 2. The combination of data sources may be configured so that the background data source is the main component. Data source i may also be included. The mixed data Zi corresponds to data before the trend change. The mixed data Zi corresponds to the first other mixed data. Multiple pieces of mixed data Zi may be generated as long as the above conditions are met.

[0038] After step SA3, the model acquisition unit 12 acquires an individual model Mi (step SA4). An individual model Mi is provided for each of one or more data sources i to be discriminated. The individual model Mi is a trained deep learning model having an input layer, a hidden layer, and an output layer. The number of hidden layers L is at least two. The deep learning model extracts features stepwise through processing in L layers. The individual model Mi is trained based on known data of each data source i to extract features of each data source i from the mixed data X, Yi, and Zi. The training method can be selected from supervised learning, semi-supervised learning, contrastive learning, unsupervised learning, and any other method. As an example, the individual model Mi is optimized based on known data of the corresponding data source i so as to maximize the degree of agreement with information contained in the corresponding data source i. Therefore, an individual model Mi is provided for each foreground data source i and optimized for each foreground data i. The individual model Mi is associated with the identifier of the data source i and stored in the storage device 2. The model acquisition unit 12 reads the individual model Mi associated with the identifier of the data source i from the storage device 2. In this way, the individual model Mi is acquired.

[0039] When step SA4 is performed, the feature calculation unit 13 calculates features Fi for each layer l (l is a natural number from 1 to L) for each of the mixed data X, Yi, and Zi using the individual model Mi acquired in step SA4. l X,Fi l Yi,Fi lFor example, the feature calculation unit 13 inputs the mixed data X to the individual model Mi, performs feature calculation for each layer l from the first layer to the Lth layer, and calculates the feature Fi l Similarly, for the mixed data Yi and Zi, feature calculations are performed for each layer l from the first layer to the Lth layer, and the feature Fi l Yi,Fi l As described above, the feature calculation unit 13 calculates the features Fi having various characteristics by utilizing the multi-layered structure of the individual model Mi. l X,Fi l Yi,Fi l Zi can be calculated.

[0040] When step SA5 is performed, the composition determination unit 14 determines the feature Fi of the shallow layer l1. l1 X and Fi l1 Based on Yi, it is determined whether or not the tendency of the data components other than the data source i included in the mixed data X has changed (step SA6). Specifically, the composition determining unit 14 uses the feature Fi calculated in step SA5 l X and Feature Fi l Among Yi, the pre-specified shallow layer l1 features Fi l1 X and Feature Fi l1 Yi is obtained. The shallow layer l1 means a layer that is closer to the input than the reference layer among the L layers. The reference layer can be arbitrarily specified, for example, L / 2 or L / 3. The number of layers in the shallow layer l1 may be one or more. The feature Fi l1 X and Fi l1 Yi tends to be dominated by feature components of data components other than data source i.

[0041] Next, the second change detection module 143 of the composition discriminator 14 detects the feature Fi l1 X and feature Fi l1 Based on Yi and Fi, l1 X and feature Fi l1An evaluation score (hereinafter referred to as a first evaluation score) for evaluating the difference between Yi is calculated, and whether or not the tendency of the data components of the foreground data source group included in the mixed data X has changed is determined based on a comparison between the first evaluation score and a threshold value (hereinafter referred to as a first threshold value). l1 X and feature Fi l1 Yi (hereinafter referred to as the first deviation) is used. If the first deviation is greater than the first threshold, the second change detection module 143 determines that the tendency of the data components other than the foreground data source i included in the mixed data X has changed, and if the first deviation is smaller than the first threshold, it determines that the tendency of the data components other than the foreground data source i included in the mixed data X has not changed.

[0042] The first type of deviation is the characteristic Fi l1 The number of X and the feature Fi l1 It can be arbitrarily selected depending on the combination with the number of Yi. l1 X and Feature Fi l1 The number of each of the features Fi can be set arbitrarily depending on the number of shallow layers 11 and the number of input data. That is, if the number of shallow layers 11, X, and Yi are all one, l1 X and Feature Fi l1 The number of Yi is 1, and if the number of layers in the shallow layer l1, X, or Yi is multiple, the feature Fi l1 X and Feature Fi l1 The number of Yi is one or more. l1 There is one X and feature Fi l1 When Yi is one, the discrepancy may be selected from Euclidean distance, Manhattan distance, cosine similarity, etc. l1 There is one X and feature Fi l1 When there are multiple Yi or feature Fi l1 There are multiple X and features Fi l1 When Yi is one, the deviation may be determined by Mahalanobis distance, a Gaussian mixture model, kernel density estimation, or the like. l1 There are multiple X and features Fi l1When there are multiple Yi, the discrepancy may be determined by Maximum Mean Discrepancy (MMD), Kullback-Leibler Divergence, or the like.

[0043] The first threshold may be a specific threshold embedded in the composition determining unit 14. Alternatively, the first threshold may be a threshold specified by a user or the like. The composition determining unit 14 may calculate the first threshold based on the known data and evaluation data used when training the individual model Mi.

[0044] When step SA6 is performed, the composition determination unit 14 determines the feature Fi of the deep layer l2. l2 X and Fi l2 Based on Zi, it is determined whether the data source i is included in the mixed data X (step SA7). Specifically, the composition determining unit 14 determines whether the data source i is included in the mixed data X based on the feature Fi calculated in step SA5. l X and Feature Fi l Among Zi, the pre-specified deep l2 features Fi l2 X and Feature Fi l2 The deep layer l2 means a layer that is closer to the output than the reference layer among the L layers. The reference layer can be arbitrarily specified, for example, L / 2 or 2L / 3. The number of layers in the deep layer l2 may be one or more. The feature Fi l2 X and Fi l2 Zi tends to be dominated by the feature components of the data components of data source i.

[0045] Next, the signal source determination module 141 of the composition determination unit 14 determines the characteristics Fi l2 X and feature Fi l2 Based on Zi and Fi, l2 X and feature Fi l2 An evaluation score (hereinafter referred to as the second evaluation score) is calculated to evaluate the difference between Zi and the foreground data source i, and whether the mixed data X includes the foreground data source i is determined based on a comparison between the second evaluation score and a threshold (hereinafter referred to as the second threshold). l2 X and feature Fil2 The deviation between Zi (hereinafter referred to as the second deviation), or the characteristic Fi l2 The residuals from the temporal and / or spatial average of X are used. The signal source identification module 141 determines that the mixed data X includes the foreground data source i if the second evaluation score is greater than the second threshold, and determines that the mixed data X does not include the foreground data source i if the second evaluation score is less than the second threshold.

[0046] The second type of deviation is the characteristic Fi l2 The number of X and the feature Fi l2 It can be arbitrarily selected depending on the combination with the number of Zi. l2 X and Feature Fi l2 The number of each of Zi can be set arbitrarily depending on the number of layers in the deep layer l2 and the number of input data. That is, if the number of layers in the deep layer l2, X, and Zi are all one, the feature Fi l2 X and Feature Fi l2 The number of Zi is 1, and if the number of layers in the deep layer l2, X, or Zi is multiple, the feature Fi l2 X and Feature Fi l2 The number of Zi is one or more. l2 There is one X and feature Fi l2 When Zi is one, the discrepancy may be selected from Euclidean distance, Manhattan distance, cosine similarity, etc. l2 There is one X and feature Fi l2 If there are multiple Zi or features Fi l2 There are multiple X and features Fi l2 When Zi is one, the deviation may be determined by Mahalanobis distance, a Gaussian mixture model, kernel density estimation, or the like. l2 There are multiple X and features Fi l2 When there are multiple Zi, the maximum average discrepancy, Kullback-Leibler divergence, etc. may be selected as the discrepancy.

[0047] FeaturesFi l2 The residuals from the temporal and / or spatial average of X are the features Fi l2Principal component analysis or Mahalanobis distance of X is used. Principal component analysis expresses the deviation from the low-dimensional space defined by the product of residuals, i.e., the covariance matrix. Mahalanobis distance expresses the Euclidean distance from the mean normalized by the covariance matrix. Note that the feature Fi l2 X may be mixed data X.

[0048] The second threshold may be a specific threshold embedded in the composition determining unit 14. Alternatively, the second threshold may be a threshold specified by a user or the like. The composition determining unit 14 may calculate the second threshold based on known data used in training the individual model Mi and a small amount of evaluation data.

[0049] When step SA7 is performed, the composition determination unit 14 determines the feature Fi of the deep layer l3. l3 X,Fi l3 Based on Yi, it is determined whether the tendency of the data component of the data source i included in the mixed data X has changed (step SA8). Step SA8 may be performed only when it is determined in step SA7 that the mixed data X includes the foreground data source i. Specifically, the composition determining unit 14 determines whether the tendency of the data component of the data source i included in the mixed data X has changed based on the feature Fi calculated in step SA5. l X and Feature Fi l Among Yi, the pre-specified deep l3 features Fi l3 X and Feature Fi l3 Yi is obtained. The deep layer l3 means a layer that is closer to the output than the reference layer among the L layers. The reference layer can be arbitrarily specified, for example, L / 2 or 2L / 3. The number of layers in the deep layer l3 may be one or more. The feature Fi l3 X and Fi l3 Yi tends to be dominated by the feature components of the data components of data source i. Deep layer l3 may be the same as deep layer l2 or may be different.

[0050] Next, the first change detection module 142 of the composition determination unit 14 detects the feature Fi l3 X and feature Fi l3 Based on Yi and Fi, l3 X and feature Fi l3An evaluation score (hereinafter referred to as a third evaluation score) for evaluating the difference between Yi is calculated, and whether or not the tendency of the data component of the foreground data source i included in the mixed data X has changed is determined based on a comparison between the third evaluation score and a threshold value (hereinafter referred to as a third threshold value). l3 X and feature Fi l3 Yi (hereinafter referred to as the third deviation degree). If the third deviation degree is greater than the third threshold, the first change detection module 142 determines that the trend of the data component of the foreground data source i included in the mixed data X has changed, and if the third deviation degree is smaller than the third threshold, it determines that the trend of the data component of the foreground data source i included in the mixed data X has not changed.

[0051] The third type of deviation is the characteristic Fi l3 The number of X and the feature Fi l3 It can be arbitrarily selected depending on the combination with the number of Yi. l3 X and Feature Fi l3 The number of each of the features Fi can be set arbitrarily depending on the number of layers in the deep layer l3 and the number of input data. That is, if the number of layers in the deep layer l3, X, and Yi are all one, l3 X and Feature Fi l3 The number of Yi is 1, and if the number of layers in the deep layer l3, X, or Zi is multiple, the feature Fi l3 X and Feature Fi l3 The number of Yi is one or more. l3 There is one X and feature Fi l3 When Yi is one, the discrepancy may be selected from Euclidean distance, Manhattan distance, cosine similarity, etc. l3 There is one X and feature Fi l3 When there are multiple Yi or feature Fi l3 There are multiple X and features Fi l3 When Yi is one, the deviation may be determined by Mahalanobis distance, a Gaussian mixture model, kernel density estimation, or the like. l3 There are multiple X and features Fi l3When there are multiple Yi, the maximum average discrepancy, Kullback-Leibler divergence, etc. may be selected as the discrepancy.

[0052] The third threshold may be a specific threshold embedded in the composition determining unit 14. Alternatively, the third threshold may be a threshold specified by a user or the like. In this case, the threshold may be calculated based on known data used in training the individual model Mi and a small amount of evaluation data.

[0053] After step SA8 is performed, the composition determining unit 14 determines whether determination has been performed for all data sources i (step SA9). If it is determined that determination has not been performed for all data sources i (step SA9: NO), the composition determining unit 14 executes steps SA2 to SA9 for the other data sources i for which determination has not been performed. The composition determining unit 14 repeats steps SA2 to SA9 until it is determined that determination has been performed for all data sources i.

[0054] If it is determined in step SA9 that the determination has been made for all data sources i (step SA9: YES), the composition determination unit 14 integrates the determination results (D4i) for step SA6 and determines whether the trend of the data components of the background data source group has changed (step SA10). As a result, a determination result (D3) indicating whether the trend of the data components of the background data source group has changed is output.

[0055] There are various methods for integrating the discrimination results (D4i). As an example, when it is determined that a trend of data components other than a specific foreground data source i has changed for a specific foreground data source i among the discrimination results (D4i) for all foreground data sources i, the integration module 144 of the composition discrimination unit 14 determines that a trend of data components of the background data source group has changed. The specific foreground data source i may be designated as the foreground data source of interest, or may be designated for all foreground data sources i. In step SA10, the data sources included in the background data source group do not need to be identified, and therefore, it is possible to equate a determination that a trend of data components other than the specific data source i has changed with a change in the trend of data components of the background data source group.

[0056] When there are multiple predetermined foreground data sources i, the integration module 144 can determine that the trend of the data components of the background data source group has changed if it is determined that the trend of any one of the data sources i has changed. As another example, when there are multiple predetermined foreground data sources i, a majority vote of the determination results may be performed. As one example, when the number of data sources determined to have changed in trend is greater than the number of data sources determined to have not changed in trend, it can be determined that the trend of the data components of the background data source group has changed.

[0057] As another example, the integration module 144 may use the evaluation scores calculated in step SA6 to determine whether the trend of the data components of the background data source group has changed. Specifically, the integration module 144 calculates an integrated score by integrating the evaluation scores corresponding to all data sources i, and determines whether the trend of the data components of the background data source group has changed based on a comparison between the integrated score and a threshold. The integrated score may be, for example, a sum or a weighted sum of the evaluation scores. If the integrated score is greater than the threshold, it may be determined that the trend of the data components of the background data source group has changed, and if the integrated score is less than the threshold, it may be determined that the trend of the data components of the background data source group has not changed.

[0058] After step SA10, the composition determination unit 14 integrates the determination results D1, D2, and D3 to generate a data composition determination result (step SA11). In step SA11, the integration module 144 of the composition determination unit 14 generates information indicating the applicability of the data composition to one of the six categories shown in FIG. 4 as the data composition determination result.

[0059] FIG. 7 shows the process of distinguishing six categories of data composition when N=1 and M=1. In the example of FIG. 7, the data source of the foreground data is a partial discharge (PD) corresponding to an abnormal signal, and the data source of the background data is a noise signal. As shown in FIG. 7, first, based on the discrimination result (D3) of step SA10, it is determined whether or not unknown noise has been detected from the mixed data (step SB1). If the trend of the data components of the background data source group has not changed, it means that the noise components contained in the mixed data are known, and if the trend of the data components of the background data source group has changed, it means that the noise components contained in the mixed data are unknown. Note that it is assumed that noise is always included in the mixed data.

[0060] If it is determined that unknown noise has not been detected (step SB1: NO), it is determined whether or not a PD data component (hereinafter referred to as a PD component) has been detected from the mixed data according to the determination result (D1) of step SA7 (step SB2). If a PD component has not been detected (step SB2: NO), the determination result of the data composition that the mixed data contains only known noise components is output.

[0061] If it is determined that a PD component has been detected from the mixed data (step SB2: YES), it is determined whether the PD component is unknown or not according to the determination result (D2) of step SA8 (step SB3). If the tendency of the PD component has not changed, it is determined that the PD component is known (step SB3: NO). In this case, the determination result of the data composition that the mixed data contains a known noise component and a known PD component is output. On the other hand, if the tendency of the PD component has changed, it is determined that the PD component is unknown (step SB3: YES). In this case, the determination result of the data composition that the mixed data contains a known noise component and an unknown PD component is output.

[0062] If it is determined that unknown noise has been detected (step SB1: YES), it is determined whether or not a PD component has been detected from the mixed data according to the determination result (D1) of step SA7 (step SB4). If a PD component has not been detected (step SB4: NO), it is determined that the mixed data contains only unknown noise components.

[0063] If it is determined that a PD component has been detected from the mixed data (step SB4: YES), it is determined whether the PD component is unknown or not according to the determination result (D2) of step SA8 (step SB5). If the trend of the PD component has not changed, it is determined that the PD component is known (step SB5: NO). In this case, the determination result of the data composition that the mixed data contains an unknown noise component and a known PD component is output. On the other hand, if the trend of the PD component has changed, it is determined that the PD component is unknown (step SB5: YES). In this case, the determination result of the data composition that the mixed data contains an unknown noise component and an unknown PD component is output.

[0064] When step SA11 is performed, the display control unit 15 displays the data composition determination result output in step SA11 on the display device 4 (step SA12). The display screen for the data composition determination result can be designed in any layout.

[0065] FIG. 8 is a diagram showing an example of a display screen I1 displaying the data composition discrimination results. The display screen I1 is displayed on the display device 4. The display screen I1 includes, for example, a first display field I11 and a second display field I12. The first display field I11 displays the mixed data X to be discriminated. The second display field I12 displays the data composition discrimination result output in step SA11. As an example, a character string such as "unknown noise + known PD" may be displayed, which indicates the data composition discrimination result corresponding to step SB5:NO in FIG. 7, that is, "The mixed data contains an unknown noise component and a known PD component." This allows the user to understand the data composition discrimination result of the mixed data X. Furthermore, by displaying the mixed data X and the data composition discrimination result side by side, the discrimination target can be visually confirmed and the reliability of the data composition discrimination result can be estimated.

[0066] This completes the data composition determination process.

[0067] 5 is an example, and the present embodiment is not limited to this. For example, the order of the mixed data X acquisition step (SA1), the mixed data Yi acquisition step (SA2), and the mixed data Zi acquisition step (SA3) is not limited to this order and may be any order. As another example, the order of the discrimination step (SA6), discrimination step (SA7), and discrimination step (SA8) is also not limited to this order and may be any order.

[0068] As another example, although features are calculated for all layers included in the individual model Mi in step SA5, features may be calculated only for the layers used in the discrimination step (SA6), discrimination step (SA7), and discrimination step (SA8).

[0069] As another example, in step SA6, the feature Fi of the shallow layer l1 l1 X,Fi l1 Based on Yi, it is determined whether there is a trend change in the data components other than the data source i. However, this embodiment is not limited to this, and the feature Fil4 X,Fi l4 Similarly, in step SA7, the determination may be made based on the deep layer l2 feature Fi l2 X,Fi l2 However, the present embodiment is not limited to this, and the presence or absence of the data source i is determined based on the feature Fi l4 X,Fi l4 Similarly, in step SA8, the determination may be made based on the deep layer l3 feature Fi l3 X,Fi l3 Yi to determine whether there is a trend change in the data components of the data source i. However, this embodiment is not limited to this, and the feature Fi l4 X,Fi l4 The determination may be made based on Yi.

[0070] As another example, in steps SA2 and SA3, Yi and Zi are generated such that the background data source is the major component other than the foreground data source i. However, this embodiment is not limited to this, and Yi and Zi may be generated such that the foreground data source j different from data source i is the major component. Such a generation method is useful when discriminating data in which another foreground data source j occurs frequently and / or is included in large amounts.

[0071] FIG. 9 is a diagram showing an evaluation of the data composition discrimination results according to this embodiment. As shown in FIG. 9, a performance index is indicated for each combination of the number of input types and noise level (NL). The number of input types in FIG. 9 refers to the number of classifications of data composition in mixed data. "Six types" refers to the six types shown in FIG. 7. "Five types" refers to those of the six types that do not distinguish between known and unknown PD components for unknown noise (in other words, do not distinguish between the presence or absence of a change in the trend of the data components). "Four types" refers to those of the six types that do not distinguish between the presence or absence of PD components for unknown noise (in other words, do not include PD components). The noise level refers to the intensity of the noise component relative to the PD component. Specifically, the noise level refers to the ratio of the maximum amplitude of the noise component to the maximum amplitude of the PD component. The comparative example is an example of random discrimination different from this embodiment. The average accuracy rate for each type of input is used as the performance index.

[0072] As shown in Fig. 9, the performance index tends to deteriorate as the number of input types and noise level increase. However, it can be seen that the method according to this embodiment improves the performance index compared to the comparative example for any number of input types. Therefore, according to this embodiment, it is possible to improve the discrimination performance of data composition.

[0073] As described above, the data composition determination process has been described with reference to FIGS. 5 to 8, but these are merely examples, and various elements can be added, deleted and / or changed.

[0074] (Variation 1) In the above embodiment, in steps SA7 and SA8, features F l2,l3 X,Fi l2,l3 Yi,Fi l2,l3 The feature calculation unit 13 according to the first modification uses the feature Fi from the shallow layer l1. l1 X,Fi l1 Yi,Fi l1 The features Fi from the deep layers l2 and l3 obtained by processing Zi and propagating it to the subsequent layers arel2,l3 X,Fi l2,l3 Yi,Fi l2,l3 Zi may be used. As an example, the feature calculation unit 13 may calculate the feature Fi l1 From X, the feature Fi l1 The residual feature is extracted by subtracting the spatial average difference of X, and the residual feature is assigned a feature Fi l1 By adding Zi, a feature (hereafter referred to as a composite feature) is generated, and by propagating the composite feature to layers after the shallow layer l1, the feature Fi from the deep layers l2 and l3 is generated. l2,l3 X,Fi l2,l3 Yi,Fi l2,l3 Zi may be calculated, and the magnitude of the residual feature may be used to determine whether foreground data source i is included.

[0075] (Variation 2) In the above embodiment, the feature calculation unit 13 performs processing independently for each data source i. However, the feature calculation unit 13 according to the second modification may perform processing common to different data sources i. As an example, the feature calculation unit 13 may perform common processing on the mixed data X, Yi, and Zi to convert them into features suitable for extraction, and input each of the mixed data X, Yi, and Zi that has undergone the common processing to a plurality of individual models Mi corresponding to the plurality of data sources i, respectively.

[0076] In addition, in the above embodiment, all the individual models Mi have the same number of layers L, but each data source i may have a different number of layers. High performance is expected by individually designing the individual models Mi according to the properties of the extracted data.

[0077] (Variation 3) In the above embodiment, the composition discriminator 14 discriminates whether or not there is a change in the trend of the data components of the background data source group in step SA10, and therefore discriminates whether or not there is a change in the trend of the data components other than the data source i in step SA6. However, this embodiment is not limited to this. The composition discriminator 14 according to the third modification discriminates whether or not there is a change in the trend of the data components of the shallow layer l1. l1 X and Fi l1Based on Yi, it may be determined whether or not the trends of the data components other than the arbitrary data source i have changed, and the determination result may be used as a determination result for whether or not the trends of the data components of the background data source group have changed. In this case, if the trends of the data components other than the arbitrary data source i have changed, it is determined that the trends of the data components of the background data source group have changed, and if the trends of the data components other than the arbitrary data source i have not changed, it is determined that the trends of the data components of the background data source group have not changed.

[0078] (Variation 4) In the above embodiment, the order in which the data source i to be identified is processed is not particularly limited. However, this embodiment is not limited to this, and the order in which the data source i to be identified is processed may be limited according to any criterion. As an example, the data sources may be processed in order starting from the data sources that tend to be included in large amounts in the current data. The order in which the data sources are processed may also be determined according to other criteria.

[0079] (Variation 5) In Modification 4, the order of processing the data source i to be discriminated is specified. Instead of specifying the order of processing the data source to be discriminated, the composition discriminator 14 according to Modification 5 may adjust the first threshold, second threshold, and third threshold used in the discrimination process (SA6, SA7, SA8) for the data source in the current iteration based on the first threshold, second threshold, and third threshold used in the discrimination process (SA6, SA7, SA8) for the other data source in the previous iteration.

[0080] (Variation 6) The composition discriminator 14 according to the sixth modification sequentially executes the processing of the signal source discrimination module 141, the first change detection module 142, and the second change detection module 143, and determines the discrimination thresholds (first threshold, second threshold, and third threshold) or evaluation scores (first evaluation score, second evaluation score, and third evaluation score) of the subsequent modules according to the discrimination results of the modules that have already been executed among the signal source discrimination module 141, the first change detection module 142, and the second change detection module 143. The order of the processing of the signal source discrimination module 141, the first change detection module 142, and the second change detection module 143 is not particularly limited, but the composition discriminator 14 according to the sixth modification executes the processing of the second change detection module 143, the signal source discrimination module 141, and the first change detection module 142 in this order.

[0081] The processing according to Variation 6 will be described in detail below. Assume that the foreground data source is an abnormal signal, and the background data source is noise. A determination of whether or not there is a trend change in the noise component, which is the dominant component among the various data components contained in the mixed data, is performed first, and the determination result is used to determine whether or not there is a subsequent abnormal signal. Specifically, to perform high-precision determination, the composition determination unit 14 switches the second threshold and / or the second evaluation score depending on whether or not there is a trend change in the noise component. As an example, if the composition determination unit 14 determines there is a trend change in the noise component, it determines whether or not there is an abnormal signal based on the residual from the average component of the features of the shallow layer l1. The individual model Mi tends to function poorly if there is a trend change in the data components of the data source during training. Especially when trained assuming a situation where the noise component is strong, the trend change in the noise component significantly affects the determination of the subsequent abnormal signal. Therefore, when a trend change in the noise component occurs, a method that is robust to trend changes, i.e., the above-mentioned method that focuses only on the residual from the average component, is preferable from the standpoint of accuracy.

[0082] As described above, the data composition determination device 100 according to the embodiment includes a mixed data acquisition unit 11, a model acquisition unit 12, a feature calculation unit 13, and a composition determination unit 14. The mixed data acquisition unit 11 acquires mixed data to be determined, which is a mixture of data components from multiple data sources. The model acquisition unit 12 acquires an individual model including multiple layers from an input layer to an output layer. The composition determination unit 14 applies the mixed data to the individual model and calculates features of the mixed data for some or all of the multiple layers. The composition determination unit 14 determines the composition of the data sources of the data components that make up the mixed data based on the calculated features.

[0083] According to the above configuration, by introducing a discrimination method that appropriately utilizes the multi-layer structure of information processing, it is possible to determine with high accuracy which data sources contain data components in mixed data and to determine changes in the trends of the included data components.In addition, because information from all layers is not used, the discrimination process is simplified, improving the maintainability of data processing, i.e., making it possible to achieve easy and stable long-term operation.

[0084] Thus, according to this embodiment, it becomes possible to determine with high accuracy the composition of data in which data components from a plurality of data sources are mixed.

[0085] Although several embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be embodied in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included within the scope and spirit of the invention, and are also included in the scope of the invention and its equivalents as defined in the claims. [Explanation of symbols]

[0086] 1...processor, 2...storage device, 3...input device, 4...display device, 5...communication device, 11...mixed data acquisition unit, 12...model acquisition unit, 13...feature calculation unit, 14...composition determination unit, 15...display control unit, 100...data composition determination device.

Claims

1. a first acquisition unit that acquires mixed data to be discriminated, the mixed data including data components from a plurality of data sources; a second acquisition unit that acquires a model including a plurality of layers from an input layer to an output layer; a calculation unit that applies the mixed data to the model and calculates features of the mixed data for each of some or all of the layers; a discrimination unit that discriminates the composition of data sources of data components that make up the mixed data based on the calculated characteristics; A data composition determination device comprising:

2. each of the plurality of data sources belongs to either a foreground data source group or a background data source group; The foreground data source group includes one or more foreground data sources, and each of the one or more foreground data sources is individually determined whether or not a data component of the foreground data source is included and whether or not a trend of the data component of the foreground data source has changed. The background data source group includes one or more background data sources, and each of the one or more background data sources is not individually determined as to whether or not it contains a data component of the background data source, but is determined as to whether or not a trend of any of the data components of the one or more background data sources has changed. the determination unit determines whether the mixed data includes each foreground data source belonging to a foreground data source group, whether a trend of the data components of the foreground data source has changed, and whether a trend of the data components of a background data source group included in the mixed data has changed; 2. The data composition determination device according to claim 1.

3. the discrimination unit includes a signal source discrimination module, a first change detection module, and a second change detection module; the signal source identification module determines, for each of the one or more foreground data sources, whether the foreground data source is included in the mixed data based on features from a first designated layer of the plurality of layers; the first change detection module determines, for each of the one or more foreground data sources, whether a trend of a data component of the foreground data source included in the mixed data has changed based on features from a second designated layer among the plurality of layers; the second change detection module determines whether a trend of the data components of the foreground data source included in the mixed data has changed based on features from a third specified layer among the plurality of layers, and integrates the determination results based on the features from the third specified layer for the one or more foreground data sources to determine whether a trend of the data components of the background data sources included in the mixed data has changed; 3. The data composition determination device according to claim 2.

4. the first acquisition unit acquires, for each of the one or more foreground data sources, first other mixed data in which a data source other than the foreground data source is dominant; the calculation unit applies the first other mixed data to the model for each of the one or more foreground data sources to calculate features of the first other mixed data from the first specified layer; The signal source identification module calculates an evaluation score for each of the one or more foreground data sources based on the characteristics of the mixed data from the first specified layer and the characteristics of the first other mixed data, and determines whether the foreground data source is included in the mixed data based on a comparison between the evaluation score and a threshold.

4. The data composition determination device according to claim 3.

5. the first designated layer is a layer that is closer to an output than a reference layer among the plurality of layers; the evaluation score is a discrepancy between the feature of the mixed data and the feature of the first other mixed data, or a residual from a temporal and / or spatial average of the feature of the mixed data; 5. The data composition determination device according to claim 4.

6. the first acquisition unit acquires, for each of the one or more foreground data sources, second other mixed data in which the foreground data source is dominant; the calculation unit applies the second other mixed data to the model for each of the one or more foreground data sources to calculate features of the second other mixed data from the second specified layer; the first change detection module calculates an evaluation score for each of the one or more foreground data sources based on the characteristics of the mixed data from the second specified layer and the characteristics of the second other mixed data, and determines whether a trend of the data component of the foreground data source included in the mixed data has changed based on a comparison between the evaluation score and a threshold; 4. The data composition determination device according to claim 3.

7. the second designated layer is a layer among the plurality of layers that is closer to an output than a reference layer, the evaluation score is a degree of discrepancy between the characteristics of the mixed data and the characteristics of the second other mixed data; 7. The data composition determination device according to claim 6.

8. the first acquisition unit acquires, for each of the one or more foreground data sources, second other mixed data in which the foreground data source is dominant; the calculation unit applies the second other mixed data to the model to calculate features of the second other mixed data from the third specified layer; the second change detection module calculates an evaluation score based on the characteristics of the mixed data from the third specified layer and the characteristics of the second other mixed data, and determines whether the tendency of the data components of the foreground data source group included in the mixed data has changed based on a comparison between the evaluation score and a threshold; 4. The data composition determination device according to claim 3.

9. the third designated layer is a layer among the plurality of layers that is closer to the input than a reference layer; the evaluation score is a degree of discrepancy between the characteristics of the mixed data and the characteristics of the second other mixed data; 9. The data composition determination device according to claim 8.

10. The determination unit sequentially executing the processes of the signal source identification module, the first change detection module, and the second change detection module; determining a discrimination threshold for a subsequent module according to a discrimination result of the module that has already been executed among the signal source discrimination module, the first change detection module, and the second change detection module; 4. The data composition determination device according to claim 3.

11. 11. The data composition determination device according to claim 10, wherein the determination unit executes the processes of the second change detection module, the signal source determination module, and the first change detection module in that order.

12. 2. The data composition determination apparatus of claim 1, wherein said model is provided for each of said one or more data sources.

13. 13. The data composition determination device according to claim 12, wherein the model is optimized based on data from a corresponding data source so as to maximize the degree of agreement with information contained in the data source.

14. 3. The data composition determination device of claim 2, wherein the determination unit determines the composition as one of the following: a first composition in which the mixed data includes only data components of a known background data source group; a second composition in which the mixed data includes data components of a known background data source group and data components of a known foreground data source; a third composition in which the mixed data includes data components of a known background data source group and data components of an unknown foreground data source; a fourth composition in which the mixed data includes only data components of an unknown background data source group; a fifth composition in which the mixed data includes data components of an unknown background data source group and data components of a known foreground data source; and a sixth composition in which the mixed data includes data components of an unknown background data source group and data components of an unknown foreground data source.

15. the foreground data source group includes one foreground data source, the group of background data sources includes one of the background data sources; 3. The data composition determination device according to claim 2.

16. the one or more foreground data sources are sources of anomalous components; the one or more background data sources are sources of noise components; 3. The data composition determination device according to claim 2.

17. the abnormal component occurs non-stationarily during a measurement period of the mixed data, the noise component occurs steadily during the measurement period; 17. The data composition determination device according to claim 16.

18. 2. The data composition determination device according to claim 1, further comprising a display control unit that displays the determined composition on a display device.

19. The processor: Acquire mixed data to be discriminated, which contains a mixture of data components from multiple data sources; Obtain a model that includes multiple layers from the input layer to the output layer, applying the mixed data to the model to calculate features of the mixed data for each of some or all of the plurality of layers; determining a data source composition of the data components that make up the mixed data based on the calculated features; A data composition determination method comprising:

20. The processor A function to acquire mixed data to be discriminated, which is a mixture of data components from multiple data sources; The ability to obtain a model containing multiple layers from the input layer to the output layer, a function of applying the mixed data to the model and calculating features of the mixed data for each of some or all of the plurality of layers; a function of determining the composition of data sources of data components constituting the mixed data based on the calculated characteristics; A data composition discrimination program that achieves this.

Images