Similar data search system, training system, and similar data search method

Abstract

To improve the accuracy of searching for similar data with minimal fluctuations. [Solution] The acquisition unit acquires a query dataset containing measurement values ​​from multiple sensors. The database stores a registered dataset containing measurement values ​​from multiple sensors. The generation unit generates an input dataset representing the difference between the query dataset and the registered dataset based on the query dataset and the registered dataset. The inference unit inputs the input dataset into a trained model and acquires an output dataset from the trained model or an intermediate output dataset which is an intermediate output of the trained model. The calculation unit calculates the similarity between the query dataset and the registered dataset based on the output dataset or the intermediate output dataset. The search unit searches the database based on the similarity.

Description

[Technical Field]

[0001] Embodiments of the present invention relate to a similar data retrieval system, a training system, and a similar data retrieval method. [Background technology]

[0002] In large-scale plants such as power plants, a large amount of process data is acquired for purposes such as monitoring the plant's performance and the health of the various systems and equipment that make up the plant. It is difficult for plant operators to constantly monitor all of this process data. For this reason, many plants are equipped with monitoring systems that use process data to detect abnormal changes in the plant.

[0003] One method for analyzing the causes of abnormal changes in current process data is to search for past process data similar to the current abnormal changes in a database. However, because fluctuations in plant process data are complex, when searching for past process data that is similar to the small fluctuations in current process data, past process data that is similar to the large fluctuations in current process data but not to the small fluctuations may also be found. [Prior art documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2017-162230 [Overview of the Initiative] [Problems that the invention aims to solve]

[0005] The problem that this invention aims to solve is to provide a similar data search system, a training system, and a similar data search method that can improve the accuracy of searching for similar data even with minute variations. [Means for solving the problem]

[0006] The similar data search system according to the embodiment includes an acquisition unit, a database, a generation unit, an inference unit, a calculation unit, and a search unit. The acquisition unit acquires a query dataset containing measurement values ​​from multiple sensors. The database stores a registered dataset containing measurement values ​​from the multiple sensors. The generation unit generates an input dataset representing the difference between the query dataset and the registered dataset based on the query dataset and the registered dataset. The inference unit inputs the input dataset into a trained model and acquires an output dataset output from the trained model or an intermediate output dataset which is an intermediate output of the trained model. The calculation unit calculates the similarity between the query dataset and the registered dataset based on the output dataset or the intermediate output dataset. The search unit searches the database based on the similarity. [Brief explanation of the drawing]

[0007] [Figure 1] This figure shows an example configuration of the similar data search system according to the first embodiment. [Figure 2] A schematic diagram showing the data collection process for registered datasets according to the first embodiment. [Figure 3] A diagram illustrating the processing procedure for searching for similar data according to the first embodiment. [Figure 4] This diagram schematically shows the processing procedure for calculating similarity according to the first embodiment. [Figure 5] This diagram schematically shows the processing procedure for generating the input dataset according to the first embodiment. [Figure 6] A diagram illustrating the data stored in the database according to the first embodiment. [Figure 7] A diagram illustrating the display screen according to the first embodiment. [Figure 8] Other diagram illustrating the display screen according to the first embodiment [Figure 9] A diagram showing an example configuration of a similar data search system related to Modification 1. [Figure 10]Figure illustrating the processing procedure of the training process of the machine learning model according to Modification Example 1 [Figure 11] Figure schematically showing the processing procedure of the training process of the machine learning model according to Modification Example 1 [Figure 12] Figure showing a configuration example of the similar data search system according to Modification Example 2 [Figure 13] Figure schematically showing the processing procedure of the generation process of the first synthetic dataset according to Modification Example 2 [Figure 14] Figure illustrating the verification result of the similar data search system according to Modification Example 2 [Figure 15] Figure showing a configuration example of the training system according to the second embodiment

Mode for Carrying Out the Invention

[0008] (First Embodiment) Hereinafter, a similar data search system, a training system, and a similar data search method according to the present embodiment will be described with reference to the drawings. Hereinafter, the term "distance" is treated as a term indicating the Euclidean distance between two data. However, the distance is not limited to the Euclidean distance. The distance according to the present embodiment is applicable even if it is, for example, the Manhattan distance, the Chebyshev distance, the Hamming distance, or the Mahalanobis distance.

[0009] FIG. 1 is a diagram showing a hardware configuration example of a similar data search system 100 according to the first embodiment. As shown in FIG. 1, the similar data search system 100 includes an information processing device 110 and a database 120. The information processing device 110 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 of the processor 1, the storage device 2, the input device 3, the display device  4, and the communication device 5 are performed via a bus. As an example, the similar data search system 100 is a system in which the information processing device 110 is an edge device such as a personal computer, and the database 120 is a server computer.

[0010] Processor 1 is an integrated circuit that controls the overall operation of the information processing device 110. For example, processor 1 includes a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), a DSP (Digital Signal Processor), and / or an FPU (Floating-Point Unit). Processor 1 may also include internal memory and I / O interfaces. Processor 1 performs various processes by interpreting and calculating programs pre-stored in the storage device 2, etc. Note that processor 1 may be partially or entirely implemented by hardware such as an ASIC (Application Specific Integrated Circuit) or FPGA (Field Programmable Gate Array).

[0011] The storage device 2 is a volatile memory and / or non-volatile memory that stores various types of data. For example, the storage device 2 stores data and settings used by the processor 1 when it performs various processes, and data generated by various processes performed by the processor 1. The storage device 2 is composed of ROM (Read Only Memory), RAM (Random Access Memory), HDD (Hard Disk Drive), SSD (Solid State Drive), integrated circuit memory, etc. The storage device 2 may also have a non-temporary computer-readable storage medium that stores programs executed by the processor 1.

[0012] Input device 3 receives various inputs from the operator. Input device 3 can include a keyboard, mouse, various switches, touchpad, touch panel display, etc. Electrical signals corresponding to the received inputs (hereinafter referred to as operation signals) are supplied to processor 1.

[0013] Display device 4 displays various data according to the control of processor 1. Display device 4 can be a CRT (Cathode-Ray Tube) display, liquid crystal display, organic EL (Electro-Luminescence) display, LED (Light-Emitting Diode) display, plasma display, or any other display as appropriate. 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 data communication with various devices connected to the information processing device 110 via a network. Operation signals may be supplied from a computer connected via the communication device 5 or from an input device on that computer, and various data may be displayed on a display device on a computer connected via the communication device 5. However, for the sake of simplicity in the following explanation, unless otherwise specified, the source of the operation signals will be the input device 3, and the destination for the display of various data will be the display device 4. The input device 3 can be replaced by a computer connected via the communication device 5 or from an input device on that computer, and the display device 4 can be replaced by a display device on a computer connected via the communication device 5, etc.

[0015] The information processing device 110 does not need to include all of the processor 1, storage device 2, input device 3, display device 4, and communication device 5. Some of the storage device 2, input device 3, display device 4, and communication device 5 may be omitted as needed. The information processing device 110 may also be provided with any additional hardware devices useful for executing the processing according to this embodiment. The information processing device 110 does not need to consist of a single physical computer; it may consist of a computer system having multiple computers connected communicably via wired or network lines. The assignment of the series of processes according to this embodiment to the multiple processors 1 implemented in each of the multiple computers can be arbitrarily configured. All processors 1 may execute all processes in parallel, or specific processes may be assigned to one or some of the processors 1, and the series of processes according to this embodiment may be executed by the entire computer system.

[0016] As shown in Figure 1, the processor 1 has a functional configuration including an acquisition unit 11, a generation unit 12, an inference unit 13, a calculation unit 14, a search unit 15, and a display control unit 16.

[0017] The acquisition unit 11 acquires various data related to similar data searches. For example, the acquisition unit 11 acquires a query dataset containing measurement values ​​from multiple sensors.

[0018] The generation unit 12 generates an input dataset representing the difference between the query dataset and the registered dataset, based on the query dataset and the registered dataset. The registered dataset is a dataset stored in the database 120.

[0019] The inference unit 13 inputs the input dataset to the trained model and obtains the output dataset from the trained model or the intermediate output dataset, which is an intermediate output of the trained model.

[0020] The calculation unit 14 calculates the similarity between the query dataset and the registered dataset based on the output dataset or the intermediate output dataset.

[0021] The search unit 15 searches the database 120 based on the similarity score.

[0022] The display control unit 16 displays various information related to similar data searches on the display device 4. For example, the display control unit 16 displays the search results of the search unit 15 on the display device 4.

[0023] Database 120 stores registered datasets containing measurement values ​​from multiple sensors.

[0024] Figure 2 shows the collection of a registered dataset. As shown in Figure 2, multivariate time series data is collected from the object being measured via multiple sensors. Multivariate time series data is time series data in which each of the sensor data output from multiple sensors is treated as one variable. For example, the multivariate time series data includes sensor data collected via N sensors, from sensor 1 to sensor N. However, the multivariate time series data may also be data that has undergone data processing such as noise reduction processing on the sensor data. As an example, the multivariate time series data consists of multiple input time series data corresponding to multiple process quantities occurring at the target facility. The target facility is a plant, such as a power plant or an industrial plant. The input time series data is, for example, time series data of the process quantities of the plant. Hereinafter, the time series data of process quantities will be referred to as process quantity data. The collected multivariate time series data is divided into multiple segments separated by predetermined time intervals. More specifically, the multivariate time series data is divided along the time series. Database 120 stores one of the divided segments as one registered dataset. Specifically, each of the multiple segments, including the divided first segment 112p and the second segment 112t, is stored in the database 120 as a registered dataset.

[0025] Furthermore, supplementary information may be associated with and stored in the registered dataset. This supplementary information may include, for example, the date and time of measurement and / or the name of the sensor. However, the supplementary information is not limited to the above. For example, supplementary conditions associated with plant quantity data may further include the operating conditions of the plant.

[0026] Hereinafter, it will be explained that database 120 stores, as a registered dataset, multiple segments of multivariate time series data, each corresponding to multiple process volumes occurring in the plant, divided into predetermined time intervals. However, database 120 is not limited to storing multiple registered datasets. Furthermore, the object of measurement is not limited to a plant. The object of measurement may be, for example, a weather phenomenon, a person, or a car. Also, the type of data in the registered dataset is not limited to process volume data. The type of data in the registered dataset may be, for example, multivariate time series data including weather data, electroencephalogram data, physical activity data, etc., or image data or video data including facial image data, fingerprint data, drive recorder data, etc.

[0027] Figure 3 shows the processing procedure for searching for a registered dataset similar to the query dataset according to the first embodiment.

[0028] As shown in Figure 3, the acquisition unit 11 acquires a query dataset (step S11). The query dataset is used as a search query when searching for a registered dataset that satisfies specific conditions from multiple registered datasets stored in the database 120. One example of a specific condition is that the similarity between the query dataset and the registered dataset is greater than or equal to a threshold. The query dataset should be of the same type of data as the registered dataset stored in the database 120. The same type of data means, for example, that it includes variables corresponding to each of the multiple variables included in the registered dataset. However, the query dataset is not limited to the same measurement target as the registered dataset. The measurement target of the query dataset can be any type of measurement target related to the registered dataset. The predetermined time should be defined to be approximately the same as the time when the registered dataset stored in the database 120 was divided. Specifically, the acquisition unit 11 acquires process amount data as a query dataset, which is the portion of the plant's process amount data from which an anomaly was detected, extracted at a predetermined time.

[0029] Figure 4 is a block diagram showing the processing flow from step S12 to step S14. The processing flow from step S12 to step S14 will be explained below using Figure 4.

[0030] When step S11 is performed, the generation unit 12 generates an input dataset 121 based on the query dataset 111 obtained in step S11 and the registered dataset 112 stored in the database 120 (step S12). For example, the input dataset 121 is the difference between the query dataset 111 and the registered dataset 112. More specifically, the generation unit 12 takes the difference between the query dataset 111 and the registered dataset 112 for each corresponding process amount data.

[0031] Figure 5 illustrates the generation process of the input dataset 121. As shown in Figure 5, the query dataset 111 and the registration dataset 112 are segments of a multivariate time series data set having N time series data points, divided into multiple segments at predetermined time intervals. The time series data points from 1 to N are different data measured by different sensors on the same time axis. Therefore, the horizontal axis of each of the N graphs included in the query dataset 111 or the registration dataset 112 is the same time axis, and the vertical axis represents the signal intensity of different process quantities. The process quantity data in the query dataset 111 and the process quantity data included in the registration dataset 112 for the same sensor number are process quantity data collected by the same sensor. The input dataset 121 is generated by taking the difference between each of the N process quantity data points in the query dataset 111 and the registration dataset 112. As a result, if the distance between the datasets is long, the fluctuations between the datasets overlap, generating a large input dataset. On the other hand, if the distance between the datasets is short, the fluctuations common to the datasets cancel each other out, generating a small input dataset. The size refers to, for example, the L2 norm. While it was explained that the input dataset 121 is generated by subtracting the registration dataset 112 from the query dataset 111, this is not the only method. For example, the input dataset 121 may be generated by subtracting the query dataset 111 from the registration dataset 112.

[0032] When step S12 is performed, the inference unit 13 inputs the input dataset 121 calculated in step S12 into the trained model 13a and obtains the output dataset 131 output from the trained model 13a (step S13). As an example, as shown in Figure 4, the trained model 13a is an autoencoder that takes the input dataset 121 as input, reduces the dimensions of the input dataset 121, and outputs an output dataset 131 which is the input dataset 121 restored to its original dimensions. The output dataset 131 is, for example, the dataset from which the input dataset 121 has been reconstructed. Specifically, the trained model 13a is an autoencoder that takes a training dataset, which is the difference between two different registered datasets 112 stored in a database 120, as input, and outputs an output dataset 13a which reconstructs the input training dataset.

[0033] More specifically, the training device inputs the training dataset into an untrained autoencoder to calculate the output dataset. If the training dataset is small, in other words, if the two registered datasets are close together, it updates the parameters of the untrained autoencoder, such as the weight parameters and bias, to minimize the error between the training dataset and the output dataset. The parameter updates are repeated until a predetermined stopping condition is met. The set of parameters at the time the predetermined stopping condition is met is assigned to the untrained autoencoder, completing the trained autoencoder. The trained model 13a is stored, for example, in the storage device 2, and processed by the information processing device 110.

[0034] By inputting the difference between the query dataset 111 and the registered dataset 112 into the trained model 13a, it is possible to obtain an output dataset 131 that takes into account small variations between the data.

[0035] Furthermore, the trained model 13a is not limited to being stored in the memory device 2. For example, the inference unit 13 may have the trained model stored in the database 120 or an external memory device processed by an external processor using cloud computing or similar methods, and the output dataset may be obtained from an external source.

[0036] When step S13 is performed, the calculation unit 14 calculates the similarity 141 between the query dataset 111 obtained in step S11 and the registered dataset 112 used to generate the input dataset in step S12, based on the output dataset 131 obtained in step S13 (step S14). The calculation unit 14 calculates the similarity 141 based, for example, on the reconstruction error between the input dataset generated in step S12 and the output dataset obtained in step S13. Specifically, the similarity 141 is the reciprocal of the magnitude of the reconstruction error. The reconstruction error is the difference between the data input to the autoencoder and the data output from the autoencoder for that input data. It is preferable to treat a smaller reconstruction error as a higher similarity 141. The calculation unit 14 calculates one similarity 141 for one input dataset 121, so one similarity 141 is calculated for one registered dataset 112.

[0037] Before executing step S15, the processes from steps S12 to S14 are performed for all the multiple registration datasets 112 stored in the database 120. Specifically, the generation unit 12 generates multiple input datasets 121 based on each of the multiple registration datasets 112 stored in the database 120 and the query dataset 111. Steps S13 and S14 are performed for each of the generated input datasets 121.

[0038] The process from step S12 to step S14 may be repeated until it is performed for all of the registered datasets 112 among the multiple registered datasets 112.

[0039] Once step S14 is performed, the database 120 stores the similarity calculated in step S14, associating it with the corresponding registered dataset (step S15).

[0040] Figure 6 illustrates the registered datasets and similarity scores stored in database 120. As shown in Figure 6, database 120 stores a list L1 in which registered datasets are linked to similarity scores calculated based on those registered datasets and supplementary information about those registered datasets. N registered datasets, ranging from 1 to N, are stored in database 120. Registered datasets are stored, for example, in the order of the time series in the multivariate time series data before splitting. Supplementary information includes information such as the measurement date and / or sensor name of the registered dataset, which is stored in advance and linked to the registered dataset. Similarity is an index that indicates the degree to which a registered dataset is similar to a query dataset. Similarity scores are calculated, for example in step S14, and then sequentially linked to the corresponding registered datasets and stored in database 120.

[0041] Furthermore, the correspondence between registered datasets and similarity scores is not limited to being stored in list format; it may be stored in any format as long as the registered datasets and similarity scores are linked. Also, the similarity scores and the correspondence between registered datasets and similarity scores are not limited to being stored in database 120. For example, the similarity scores and the correspondence between registered datasets and similarity scores may be stored in storage device 2 or in an external storage device, as long as they are readable by processor 1.

[0042] When step S15 is performed, the search unit 15 generates search results based on the similarity calculated in step S14 (step S16). As an example, the search unit 15 generates search results as a list in which some or all of the multiple registered datasets stored in the database 120 are arranged in order of the relative similarity levels. Specifically, the search results are in the form of a list in which the similarity levels of the multiple registered datasets that exceed a predetermined threshold are arranged in descending order. The predetermined threshold may be a statistical value such as the median or mean based on all the similarity levels calculated in step S14, or it may be a value arbitrarily determined by the user. Note that the search results are not limited to the above format. For example, the search unit 15 may generate search results that include only the registered dataset with the highest similarity level.

[0043] When step S16 is performed, the display control unit 16 displays the search results generated in step S16 (step S17).

[0044] Figure 7 illustrates a display screen I1 that shows search results. Display screen I1 in Figure 7 includes a display field I11. Display field I11 displays a list that combines the identification information, similarity score, and supplementary information of the registered dataset. The list is sorted in descending order of similarity. The identification information is information used to identify each of the multiple registered datasets stored in database 120. The supplementary information is information such as the measurement date and / or sensor name that is stored in association with the registered dataset stored in database 120. The similarity score is an indicator that the registered dataset and the query dataset are more similar the higher the score is in the list. The higher a registered dataset is positioned in the list, the more similar it is to the query dataset. The rows in the list may be displayed in a selectable format, for example. By displaying search results based on similarity, it is possible to easily identify registered datasets that are similar to the query dataset from among multiple registered datasets.

[0045] Figure 8 is another diagram illustrating the display screen I1 that displays search results. As shown in Figure 8, display screen I1 displays display fields I11 and I12. Display field I11 displays a list that combines the identification information, similarity, and supplementary information of the registered dataset. The list is sorted in descending order of similarity. The rows in the list displayed in display field I11 are displayed as selectable, for example. In Figure 8, the row with the string "ID52" as the identification information is selected. Selected rows are highlighted, for example, by having a thick border around the row. When a row is selected in the list displayed in display field I11, display field I12 displays information about the registered dataset associated with the selected row. Display field I12 displays, for example, the time-series data of the registered dataset shown in the selected row and the input dataset based on that registered dataset. Display field I12 is displayed overlaid on the list so as not to cover the selected row. By displaying the registration dataset shown in the selected list row and the input dataset based on that registration dataset, users can easily determine whether the registration dataset displayed in the list is the desired registration dataset.

[0046] When step S17 is performed, the search process for registered datasets similar to the query dataset according to the first embodiment is terminated.

[0047] Furthermore, this embodiment is applicable even if there is only one registered data stored in the database 120. For example, the search unit 15 generates search results by classifying registered data sets above a predetermined threshold as similar registered data sets to the query data set, and by classifying registered data sets below the predetermined threshold as not similar registered data sets to the query data set. Also, the similarity score 141 is not limited to the above. For example, the similarity score 141 may be the reconstruction error between the input data set 121 and the output data set 131 itself.

[0048] In this first embodiment, an input dataset representing the difference between the query dataset and the registered dataset is input to a trained model, and by using the similarity based on the output dataset output from the trained model, it is possible to search for similar data not only based on close distances between data but also on minute variations. Consequently, it is possible to improve the search accuracy of similar registered datasets based on minute variations in the query dataset. Furthermore, by displaying search results using similarity based on reconstruction error, users can easily identify data similar to the query data.

[0049] Furthermore, by applying this embodiment to the retrieval of plant quantity data related to monitoring and controlling the plant's operating status, the following effects can be achieved: Access to past trouble records from measurement dates and sensor names linked to the retrieved past process quantity data makes it possible to quickly perform cause analysis and, consequently, support for countermeasures when abnormalities occur. In addition, by constantly retrieving past process quantity data similar to the current process quantity data, it is possible to monitor the degree of similarity at any time and determine the current health, thereby supporting health evaluation. Moreover, it is possible to access operating records linked to the retrieved past process quantity data, and consequently, by quickly grasping the operating information at the time, which is used to determine current operating conditions based on past plant operation information, it is possible to support the setting of operating conditions.

[0050] (Variation 1) The similar data search system according to Modification 1 further comprises a training unit 17.

[0051] Figure 9 shows an example of the hardware configuration of the similar data search system 100 according to Modification 1. As shown in Figure 9, the similar data search system 100 has an information processing device 110 and a database 120. The information processing device 110 is a computer having a processor 1, a storage device 2, an input device 3, a display device 4, and a communication device 5. The transmission and reception of data and various signals between the processor 1, storage device 2, input device 3, display device 4, and communication device 5 is performed via a bus. As an example, the similar data search system 100 is a system in which the information processing device 110 is an edge device such as a personal computer, and the database 120 is a server computer.

[0052] Database 120 stores multiple registered datasets, each containing measurements from multiple sensors.

[0053] As shown in Figure 9, the processor 1 has a functional configuration including an acquisition unit 11, a generation unit 12, an inference unit 13, a calculation unit 14, a search unit 15, a display control unit 16, and a training unit 17.

[0054] The acquisition unit 11 acquires two different registered datasets from among multiple registered datasets.

[0055] The generation unit 12 generates a training dataset that represents the difference between two different registration datasets from among multiple registration datasets, based on two different registration datasets from among multiple registration datasets.

[0056] The training unit 17 trains a machine learning model for use in the inference unit 13. For example, the training unit 17 trains the machine learning model to take a training dataset as input and output an output dataset for the input training dataset.

[0057] Figure 10 shows the processing procedure for training the machine learning model according to Modification Example 1. Figure 11 is a schematic diagram showing the flow of the training process according to Modification Example 1. The following explanation will use Figures 10 and 11.

[0058] As shown in Figure 10, the acquisition unit 11 acquires two different registered datasets from among the multiple registered datasets stored in the database 120 (step S21). For example, when the acquisition unit 11 performs the training process of an untrained machine learning model 17a, it acquires combinations of the first registered dataset 112a and the second registered dataset 112b multiple times, according to the pre-set batch size of the training data and the number of epochs of the training process. More specifically, the acquisition unit 11 acquires combinations of two different registered datasets from among the multiple registered datasets without duplication for all of the multiple registered datasets.

[0059] Specifically, if there are N registered datasets, numbered from 1 to N, stored in the database 120, then if the first registered dataset 112a is registered dataset 1, then there are N-1 possible second registered datasets 112b, ranging from registered dataset 2 to registered dataset N. Since there are several ways to select the first registered dataset, from registered dataset 1 to registered dataset N, the acquisition unit 11 acquires N(N-1) combinations of two registered datasets. The similar data search device according to Modification Example 1 can improve training accuracy by training an untrained machine learning model 17a on all registered datasets stored in the database 120.

[0060] However, as the number of registered datasets stored in database 120 increases, the number of combinations of two registered datasets to retrieve becomes O(N). 2Because the amount of data increases, the training process may not be completed in a realistic amount of time due to constraints such as the memory of the information processing device 110 and the processing speed of the processor 1. Incidentally, for example, during steady operation of power plants including thermal power plants, a large number of process quantity data with approximately the same fluctuations are included in multiple registered datasets. In multiple registered datasets containing process quantity data with approximately the same fluctuations, increasing the number of combinations to acquire does not increase the amount of information that is significant for training, and the training accuracy does not improve significantly, so the significance of generating training datasets for all combinations of multiple registered datasets stored in the database 120 is reduced.

[0061] In such cases, instead of acquiring two registration datasets for all combinations of registration datasets, the process of acquiring two different registration datasets randomly selected from multiple registration datasets stored in the database 120 is performed a predetermined number of times. The predetermined number of times is determined based on the time required for the generation unit 12 to generate the training dataset. For example, the predetermined number of times should be determined so that the generation unit 12 can complete the process of generating the training dataset in a realistic amount of time. By acquiring two different registration datasets randomly selected from multiple registration datasets stored in the database 120 a predetermined number of times, it is possible to reduce the number of combinations of the first registration data 112a and the second registration data 112b acquired by the acquisition unit 11 without reducing the training accuracy.

[0062] When step S21 is performed, the generation unit 12 generates a training dataset as the difference between the two registered datasets acquired in step S21 (step S22). When a first registered dataset 112a and a second registered dataset 112b that are close in distance are acquired, the first registered dataset 112a and the second registered dataset 112b will often have similar operating conditions in terms of plant process quantity data. Specifically, the first registered dataset 112a and the second registered dataset 112b can be treated as plant quantity data in which common fluctuations and small fluctuations between the two registered datasets are combined.

[0063] In plant process volume data, the common fluctuations between the first registered dataset 112a and the second registered dataset 112b are complex, and in similar data searches, the small fluctuations that we want to focus on may be buried in the common fluctuations, and the desired similar data may not be found. When the distance between the data of the first registered dataset 112a and the second registered dataset 112b is small, the training dataset, which is the difference, cancels out the common fluctuations between the first registered data 112a and the second registered dataset 112b, and represents the difference between the first small fluctuation of the first registered dataset 112a and the second small fluctuation of the registered dataset 112b. By performing the difference calculation between the first registered dataset 112a and the second registered dataset 112b, the generation unit 12 can generate a training dataset that represents the difference between the first registered dataset 112a and the second registered dataset 112b.

[0064] When step S22 is performed, the training unit 17 inputs the training dataset generated in step S22 into the untrained machine learning model 17a and obtains the reconstructed dataset output by the untrained machine learning model 17a (step S23). The machine learning model 17a is, for example, an autoencoder. By training the untrained machine learning model based on the training dataset, the machine learning model 17a can learn about a first minute variation and / or a second minute variation from which the variation common to the two registered datasets has been excluded.

[0065] When step S23 is performed, the training unit 17 calculates a loss 17b based on the training dataset input to the untrained machine learning model 17a in step S22 and the reconstruction error dataset calculated in step S23 (step S24). The features that the machine learning model 17a learns are described below.

[0066] When the distance between the first registration data 112a and the second registration data 112b is small, the plant at the time of acquisition of the first registration data 112a and the plant at the time of acquisition of the second registration data 112b are often in similar operating conditions. Registration datasets acquired under similar operating conditions have similar characteristics of small fluctuations. On the other hand, in a small number of cases, even if the distance between the first registration data 112a and the second registration data 112b is small, the characteristics of small fluctuations between these registration datasets may differ. In this case, the plant at the time of acquisition of the first registration data 112a and the plant at the time of acquisition of the second registration data 112b are not in similar operating conditions. Since the training process is a statistical process, the untrained machine learning model 17a is trained based on the data characteristics dominated by the number of training datasets, and the characteristics of a small number of data do not have a significant impact on the training of the untrained machine learning model 17a. That is, the untrained machine learning model 17a is trained on the characteristics of small fluctuations between process quantity data that are present in similar operating conditions.

[0067] An untrained machine learning model 17a can take on various network configurations, but training may become unstable if the configuration does not have a bottleneck structure. For this reason, the following describes the details of loss 17b when the untrained machine learning model 17a is used as an autoencoder that reduces the dimensionality of the input data and then outputs data restored to the same dimensionality as the input data.

[0068] The autoencoder is trained such that the closer the distance between the first registered data 112a and the second registered data 112b is, the smaller the reconstruction error becomes. In the above training process, for features in a small number of cases where the operating states of the plants are not similar even though the distance is short, it can be regarded as being statistically excluded from the training. To realize such training, the loss L, with the batch size being N, the training dataset being X k , the output data being X k ′ (1 ≤ k ≤ N), and S(X k ) being ||X k || 2 , and δ being an arbitrary constant for preventing division by zero, is expressed by the following formula (1).

[0069]

Equation

[0070] Also, the following formula (2), which is a developed form of formula (1), may be used as the loss L.

[0071]

Equation

[0072]

number

[0073] When step S24 is performed, the training unit 17 updates the parameters of the untrained machine learning model 17a to minimize the loss 17b calculated in step S24 (step S25). The parameter update is performed, for example, by the update circuit 17c. Based on the loss 17b calculated in step S24, the update circuit 17c updates the parameters of the untrained machine learning model 17a using a deep learning parameter optimization algorithm such as Adam or SGD.

[0074] The process in steps S21–S25 is repeated while changing the combination of the two registered datasets until the parameter update termination condition is met. For example, the update termination condition can be set to any condition such as the number of iterations of steps S21–S25 reaching a predetermined number, or the loss being less than a predetermined value. When the update termination condition is met, the training process of the machine learning model according to the modified example 1 is terminated.

[0075] The registered dataset used for training is not limited to the datasets stored in database 120. For example, the training dataset may be generated based on a dataset stored in an external storage device, or based on a dataset stored in a portable storage medium.

[0076] The following describes the case where a trained autoencoder, trained by the training process described in Modification 1, is used in the inference unit 13.

[0077] This section describes the case where the distance between the query dataset and the first registration data is equal to the distance between the query dataset and the second registration dataset. In this case, the L2 norm of the first input dataset, which is the difference between the query dataset and the first registration data, is equal to the L2 norm of the second input dataset, which is the difference between the query dataset and the second registration data. The trained autoencoder has learned the features of the training dataset, which is the difference between two different registration datasets among the multiple registration datasets stored in the database 120. Therefore, if the features of the first input dataset are more similar to the features learned by the trained autoencoder than the features of the second input dataset, the first reconstruction error, which is the difference between the first input dataset and the first output dataset, will be smaller than the second reconstruction error, which is the difference between the second input dataset and the second output dataset. As a result, the first reconstruction error will be smaller than the second reconstruction error.

[0078] Therefore, when the distance between the query dataset and the registered dataset is small, the registered dataset is not considered similar to the query dataset. Rather, in addition to the distance, the registered dataset that has similar characteristics in terms of minute fluctuations in process volume data under similar operating conditions is considered similar to the query dataset.

[0079] Furthermore, the training unit 17 may train an untrained machine learning model using the features of the registered dataset. Since features are data organized to represent the characteristics of the original process data, it is possible to improve the training accuracy of the untrained machine learning model.

[0080] According to Modification 1, it is possible to generate a machine learning model trained by focusing on minute variations that exclude common variations between two registered datasets. Furthermore, by using the machine learning model trained in the training unit 17 in the inference unit 13, it is possible to improve the search accuracy of registered datasets with similar minute variations in the query dataset.

[0081] (Modification 2) The similar data search system according to the modified example 2 further comprises a preprocessing unit 18.

[0082] Figure 12 shows an example of the hardware configuration of the similar data search system 100 according to Modification 2. As shown in Figure 12, the similar data search system 100 has an information processing device 110 and a database 120. The information processing device 110 is a computer having a processor 1, a storage device 2, an input device 3, a display device 4, and a communication device 5. The transmission and reception of data and various signals between the processor 1, storage device 2, input device 3, display device 4, and communication device 5 is performed via a bus. As an example, the similar data search system 100 is a system in which the information processing device 110 is an edge device such as a personal computer, and the database 120 is a server computer.

[0083] As shown in Figure 12, the processor 1 has a functional configuration including an acquisition unit 11, a generation unit 12, an inference unit 13, a calculation unit 14, a search unit 15, a display control unit 16, and a preprocessing unit 18.

[0084] The preprocessor 18 includes a first autoencoder and a second autoencoder. The first autoencoder performs dimensionality reduction and restoration of the query dataset and the registration dataset. Specifically, the first autoencoder takes the query dataset as input as the first intermediate dataset, reduces the dimensions of the input first intermediate dataset, and outputs either a first reconstructed dataset, which is the first intermediate dataset with reduced dimensions, restored to the same dimensions as the input first intermediate dataset, or a first feature dataset, which is the first intermediate dataset with reduced dimensions. The first autoencoder also takes the registration dataset as input as the first intermediate dataset and outputs the first reconstructed dataset and the first feature dataset.

[0085] The second autoencoder is a different autoencoder from the first autoencoder, performing dimensionality reduction and restoration of the query dataset and registration dataset. Specifically, the second autoencoder receives the second intermediate dataset as input, which is the difference between the first intermediate dataset and the first reconstructed dataset output by the first autoencoder for the input of the first intermediate dataset. The second autoencoder reduces the dimensionality of the input second intermediate dataset and outputs a second feature dataset, which is the second intermediate dataset with reduced dimensionality.

[0086] The generation unit 12 generates an input dataset based on a first feature dataset based on a query dataset and a first feature dataset based on a registration dataset, or based on a second feature dataset based on a query dataset and a second feature dataset based on a registration dataset.

[0087] Database 120 stores a first feature dataset and a second intermediate output dataset based on the registered dataset.

[0088] Figure 13 is a block diagram illustrating the flow of the generation process for the first preprocessed dataset of the preprocessing unit 18 according to Modification 2. The generation process for the first preprocessed dataset is preferably performed between steps S11 and S12 in Figure 3. As shown in Figure 13, the query dataset 111 acquired in step S11 as the first intermediate dataset is input to the trained machine learning model 18a and the first difference circuit 18b. The trained machine learning model 18a is, for example, the first autoencoder. Hereinafter, the trained machine learning model 18a will be referred to as the first autoencoder 18a. The first autoencoder 18a outputs a first reconstructed dataset by reconstructing the input query dataset 111. The first reconstructed dataset is a dataset that reproduces the rough fluctuations of multivariate time series data that are expected to be acquired during normal plant operation. Rough fluctuations refer to, for example, the low-frequency components when the time series data is represented by a combination of high-frequency and low-frequency components. The first reconstructed dataset output from the first autoencoder 18a is input to the difference circuit 18b.

[0089] The difference circuit 18b outputs a second intermediate dataset, which is the difference between the input query dataset 111 and the first reconstructed dataset. Since the second intermediate dataset is the difference between the query data 111 and the first reconstructed dataset, the large fluctuations under normal conditions of the multivariate time series data are reduced, and the dataset containing minute fluctuations is extracted. The second intermediate dataset output from the first difference circuit is input to the trained machine learning model 18c. The trained machine learning model 18c is, for example, a second autoencoder. Hereafter, the trained machine learning model 18c will be referred to as the second autoencoder 18c.

[0090] The second autoencoder 18c outputs a second feature dataset obtained by reducing the dimensionality of the input second intermediate dataset. The second feature dataset is a dataset in which features relating to subtle fluctuations in multivariate time series data have been extracted. Subtle fluctuations refer to, for example, the high-frequency components when the time series data is represented as a combination of high-frequency and low-frequency components. The preprocessing unit 18 outputs the second feature dataset as the first preprocessed dataset 181.

[0091] Note that the dataset output from the preprocessing unit 18 is not limited to the second feature dataset. For example, the dataset output from the preprocessing unit 18 may be the first feature dataset. The first feature dataset is a dataset in which the first intermediate dataset is input to the first autoencoder 18a, and features related to the rough fluctuations of the output multivariate time series data are extracted. The preprocessing unit 18 may also appropriately select either the first or second feature dataset according to the user's selection via the input device 3, and output only the selected dataset as the first preprocessed dataset 181. In this case, datasets that are not selected do not need to be calculated. For example, if the first feature dataset is selected, the first autoencoder 18a does not need to output the first reconstructed dataset, and the second autoencoder 18c does not need to output the second feature dataset. If the second feature dataset is selected, the first autoencoder 18a does not need to output the first feature dataset. Based on the user's selection, the first preprocessing dataset 181 is output as either a first or second feature dataset, allowing for the use of broad or subtle variations in similar data searches. Furthermore, if the first feature dataset is selected as the first preprocessing dataset 181, the second feature dataset is not calculated, thereby reducing the time required for similar data searches.

[0092] Furthermore, the modified form 2 is also applicable in cases where the first autoencoder 18a or the second autoencoder 18c does not output a reconstructed dataset, and only the decoder portion of the autoencoder trained as the first autoencoder 18a or the second autoencoder 18c is used.

[0093] Furthermore, the data input to the preprocessing unit 18 is not limited to raw data. For example, the query data 111 may be a dataset based on data from which some of the time series data from multivariate time series data has been extracted.

[0094] By applying the processing procedure shown in Figure 13 to the registered dataset, the preprocessing unit 18 outputs a first feature dataset and a second feature dataset based on the registered dataset as a second preprocessing dataset. The second preprocessing dataset is generated by the preprocessing unit 18 and stored in the database 120 at the same time that the registered dataset is stored in the database 120. At this time, it is desirable that a first feature dataset and a second feature dataset are generated for each registered dataset. By having the second preprocessing dataset stored in the database 120 in advance, it is possible to reduce the time required to generate the second preprocessing dataset during similar data searches.

[0095] The generation unit 12 outputs an input dataset by taking the difference between the first preprocessing dataset 181 and the second preprocessing dataset. For example, the generation unit 12 outputs the difference between the first preprocessing dataset and the second preprocessing dataset selected by the user as the input dataset. Specifically, if the first preprocessing dataset is the first feature dataset, the generation unit 12 uses only the second preprocessing dataset, which is the first feature dataset, to generate the input dataset. Also, if the first preprocessing dataset is the second feature dataset, the generation unit 12 uses only the second preprocessing dataset, which is the second feature dataset, to generate the input dataset. The above process corresponds to step S12 in Figure 3. After the above process, the same processing as in step S13 and onward in Figure 3 can be applied.

[0096] Figure 14 shows the verification results using power plant operation data. As shown in Figure 14, the search accuracy when searching for similar data using four methods for each of the four cases is shown in a 5x6 table T1. The first row of T1 contains six strings in six cells, from left to right: “Case,” “Number of Occurrences,” “Nearest Neighbor Method,” “Nearest Neighbor Method + (1),” “Nearest Neighbor Method + (1) + (2),” and “This Method + (1) + (2).” A case represents the operating state of the plant. The number of occurrences represents the number of times the case occurred. Since the similar data search process for the case is executed each time the case occurs, the number of occurrences is equal to the number of times the similar data search process using the case as the query dataset is performed. Search accuracy is the ratio of the number of times a dataset similar to the query dataset was found to the number of similar data search processes. Search accuracy may also be called, for example, the accuracy rate. The nearest neighbor method represents the general k-nearest neighbor method. This method represents the similar data search method using the similar data search system according to the first embodiment. (1) indicates that a dataset based on data from which some time series data from multivariate time series data has been extracted was used as the query dataset and registration dataset. (2) indicates that low-dimensional features output from the first or second autoencoder were used. In other words, Method + (1) + (2) represents a similar data search method using the similar data search system related to Modification 2. The plant quantity data targeted was validated on 5 years of data for approximately 300 process quantities of a power plant. The weighted average is the value obtained by dividing the sum of values ​​where the number of occurrences for each case is weighted by the search accuracy by the sum of the number of occurrences for each case in each method. According to Table T1, in Case 1, Case 3 and the weighted average, the similar data search system related to Modification 2 is shown to have higher search accuracy than the nearest neighbor method, nearest neighbor method + (1), and nearest neighbor method + (1) + (2).

[0097] According to Modification 2, it is possible to extract abnormal plant quantity data from plant quantity data containing complex components. Furthermore, it is possible to improve the accuracy of searching for similar data in power plants.

[0098] (Variation 3) The machine learning model described in Modification 3 is applicable to machine learning models other than autoencoders. For example, this embodiment is applicable to machine learning models such as multi-layer perceptrons (MLPs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs). These should be used appropriately depending on the type of data in the input query dataset.

[0099] According to Modification 3, the similar data search system can also be applied to multivariate time-series data including weather data, electroencephalogram data, and physical activity data, as well as image and / or video data including facial photographs, fingerprint data, and dashcam footage, in addition to plant quantity data.

[0100] (Modification 4) The calculation unit 14 in the modified example 4 calculates similarity based on an intermediate output dataset, which is an intermediate output of the trained model. The intermediate output dataset is, for example, a set of features from which the dimensionality of the input dataset has been reduced. Specifically, the calculation unit 14 may calculate similarity based on the features from which the dimensionality has been reduced by the autoencoder used in the inference unit 13. More specifically, the calculation unit 14 calculates the L2 norm of the features from which the dimensionality has been reduced by the autoencoder as the similarity. However, the intermediate output dataset is not limited to features from which the dimensionality has been reduced by the autoencoder; this embodiment can also be applied to any state variables in the trained model.

[0101] According to Modification 4, by calculating the similarity based on the intermediate output data, which is the intermediate output of the trained model, it is possible to calculate the similarity even when it is not possible to calculate the difference.

[0102] (Second Embodiment) The first embodiment is treated as a similar data search system. The second embodiment is a training system for training a machine learning model of the similar data search system according to the first embodiment. The training system according to the third embodiment will be described below. However, components having the same function as those in the first embodiment are denoted by the same reference numerals and described redundantly only when necessary.

[0103] Figure 15 shows an example configuration of the training system 200 according to the second embodiment. As shown in Figure 14, the training system 200 has an information processing device 210 and a database 120. The information processing device 210 is a computer having a processor 1, a storage device 2, an input device 3, a display device 4, and a communication device 5. Data and various signals from the processor 1, storage device 2, input device 3, display device 4, and communication device 5 are transmitted and received via a bus. As an example, the training system 200 is a system in which the information processing device 210 is an edge device such as a personal computer and the database 120 is a server computer.

[0104] As shown in Figure 15, the processor 1 has a functional configuration including an acquisition unit 21, a generation unit 22, and a training unit 23.

[0105] The acquisition unit 21 corresponds to the acquisition unit 11 in Modification Example 1. The generation unit 22 corresponds to the generation unit 12 in Modification Example 1. The training unit 23 corresponds to the training unit 17 in Modification Example 1.

[0106] According to the second embodiment, it is possible to generate a trained model for use in similar data search without having a function related to similar data search.

[0107] Thus, according to some of the embodiments described above, it becomes possible to provide a similar data search system, a training system, and a similar data search method that can improve the accuracy of searching for similar data with minute variations.

[0108] While several embodiments of the present invention have been described, these embodiments are presented as examples only and are not intended to limit the scope of the invention. These novel embodiments can be carried out in a variety of other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their variations are included in the scope and spirit of the invention, as well as in the claims of the invention and its equivalents. [Explanation of Symbols]

[0109] 1...Processor, 2...Storage device, 3...Input device, 4...Display device, 5...Communication device, 11...Acquisition unit, 12...Generation unit, 13...Inference unit, 14...Calculation unit, 15...Search unit, 16...Display control unit, 100...Similar data search system, 110...Information processing device, 120...Database

Claims

1. An acquisition unit that acquires a query dataset containing measurement values ​​from multiple sensors, A database that stores a registered dataset containing measurement values ​​from the aforementioned multiple sensors, A generation unit generates an input dataset representing the difference between the query dataset and the registration dataset based on the query dataset and the registration dataset, An inference unit that inputs the input dataset to a trained model and obtains an output dataset from the trained model or an intermediate output dataset which is an intermediate output of the trained model, A calculation unit that calculates the similarity between the query dataset and the registered dataset based on the output dataset or the intermediate output dataset, A search unit that searches the database based on the similarity, A similar data search system equipped with the following features.

2. The similar data search system according to claim 1, wherein the trained model is an autoencoder that takes the input dataset as input, reduces the dimension of the input dataset, and outputs an output dataset restored to the dimension of the input dataset.

3. The similarity search system according to claim 2, wherein the calculation unit calculates the similarity based on the reconstruction error between the input dataset and the output dataset.

4. The similar data search system according to claim 1, wherein the query dataset and the registration dataset are one of a plurality of segments into which multivariate time series data is divided at predetermined time intervals, or features extracted from one of the segments.

5. The similar data search system according to claim 4, wherein the multivariate time series data is a plurality of input time series data corresponding to a plurality of process quantities occurring in a target facility including a plant.

6. The similar data search system according to claim 1, wherein the generation unit generates the difference between the query dataset and the registered dataset as the input dataset.

7. The similar data search system according to claim 1, wherein the query dataset and the registration dataset include multivariate time series data including weather data, electroencephalogram data, and physical activity data, facial image data, fingerprint data, and image and / or video data including drive recorder data.

8. The similarity search system according to claim 1, wherein the similarity is the L2 norm of the output dataset or the intermediate output dataset.

9. The similar data search system according to claim 1, further comprising a display unit that displays search results for registered datasets similar to the query dataset.

10. The database stores the registered dataset along with supplementary information including the measurement date and / or the name of the sensor. The similar data search system according to claim 9, wherein the display unit displays the supplementary information linked to the registered dataset included in the search results together with the search results.

11. The database stores multiple registered datasets, each containing measurement values ​​from the multiple sensors. The similar data search system according to claim 1, wherein the generation unit generates the input dataset for each of the multiple registration datasets included in the plurality of registration datasets.

12. The database stores multiple registered datasets, each containing measurement values ​​from the multiple sensors. The similarity data search system according to claim 1, wherein the search unit generates search results in which some or all of the plurality of registered datasets are arranged in order of the relative similarity between them.

13. The preprocessor further comprises a first autoencoder and a second autoencoder different from the first autoencoder, The first autoencoder takes the query dataset as input as a first intermediate dataset, reduces the dimensionality of the input first intermediate dataset, and restores the dimensionally reduced first intermediate dataset to a dataset with the same dimensionality as the input first intermediate dataset, thereby obtaining a first reconstructed dataset or a first feature dataset which is the dimensionally reduced first intermediate dataset; and takes the registration dataset as input as a first intermediate dataset, and outputs the first reconstructed dataset and the first feature dataset. The second autoencoder receives a second intermediate dataset as input, which is the difference between the first intermediate dataset and the first reconstructed dataset or the first feature dataset output by the first autoencoder in response to the input of the first intermediate dataset. The second autoencoder reduces the dimensionality of the input second intermediate dataset and outputs a second feature dataset, which is the second intermediate dataset with reduced dimensionality. The database stores the first feature dataset and the second feature dataset based on the registered dataset. The similar data search system according to claim 1, wherein the generation unit generates the input dataset based on the first feature dataset based on the query dataset and the first feature dataset based on the registration dataset, or based on the second feature dataset based on the query dataset and the second feature dataset based on the registration dataset.

14. It further includes a training unit for training machine learning models to be used in the inference unit, The database stores multiple registered datasets, each containing measurement values ​​from the multiple sensors. The acquisition unit acquires two different registration datasets from the plurality of registration datasets, The generation unit generates a training dataset representing the difference between the two registration datasets based on the two registration datasets. The similar data search system according to claim 1, wherein the training dataset is input and the machine learning model is trained to output the output dataset for the input training dataset.

15. A database that stores multiple registered datasets containing measurements from multiple sensors, An acquisition unit that acquires two different registration datasets from the aforementioned plurality of registration datasets, A generation unit generates a training dataset that represents the difference between the two registration datasets based on the two registration datasets, A training unit that takes the aforementioned training dataset as input and trains a machine learning model to output an output dataset for the input training dataset, A training system equipped with the following features.

16. The aforementioned machine learning model is an autoencoder, The training system according to claim 15, wherein the training unit updates the parameters of the machine learning model to minimize the loss based on the training dataset input to the machine learning model and the reconstruction error, which is the difference between the training dataset and the output dataset output by the machine learning model with respect to the input training dataset.

17. The training system according to claim 16, wherein the loss has a term that correlates with the magnitude of the reconstruction error when the size of the training dataset is less than 1.

18. The training system according to claim 15, wherein the generation unit performs the process of generating the training dataset from two randomly selected registration datasets from the plurality of registration datasets a predetermined number of times, excluding combinations of the two registration datasets.

19. The training system according to claim 18, wherein the predetermined number of times is determined based on the time required for the generation unit to process the training dataset.

20. Computers Obtain a query dataset containing measurements from multiple sensors, A registered dataset containing the measurement values ​​of the aforementioned multiple sensors is stored in a database. Based on the query dataset and the registration dataset, an input dataset representing the difference between the query dataset and the registration dataset is generated. The input dataset is input to the trained model, and the output dataset output from the trained model or the intermediate output dataset, which is an intermediate output of the trained model, is obtained. Based on the output dataset or the intermediate output dataset, the similarity between the query dataset and the registration dataset is calculated. Based on the similarity, search the database. A similar data search method that includes the following features.

Images