Data compression device, data compression method, and data compression program

Abstract

To facilitate accurate learning of a model.SOLUTION: The data compression apparatus 100 acquires target data for learning from predetermined time-series data. The data compression apparatus 100 constructs subspaces by dividing the data for learning based on the explanatory variables included in the acquired data for learning. The data compression device 100 compresses data for learning for each constructed subspace to generate compressed data for learning to be used for learning a regression model.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to a data compression device, a data compression method, and a data compression program. [Background technology]

[0002] Machine learning models such as neural networks, including fuzzy neural networks, are sometimes trained to learn the relationships between time-series data collected by sensors installed in chemical plants, etc., and the trained models are then used to detect abnormalities and continuously calculate characteristic values ​​that are difficult to measure.

[0003] For example, in the case of anomaly sign detection, learning is performed to build a model for detecting anomalies. Therefore, a technique for efficiently creating an anomaly detection model by calculating an evaluation value of the anomaly detection model is known as a conventional technique (see, for example, Patent Document 1). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2022-056746 Summary of the Invention [Problem to be solved by the invention]

[0005] However, with the above-mentioned conventional techniques, it may be difficult to accurately train a model. For example, the conventional technique is a method of constructing a nonlinear regression model and using it for anomaly detection. However, with the conventional technique, data obtained from a sensor or the like is used as is when constructing the nonlinear regression model, so it is difficult to accurately train the model when time-series data with a bias in data distribution is used as training data. [Means for solving the problem]

[0006] Therefore, in order to solve the above-mentioned problems and achieve the object, the data compression device of the present invention is characterized by having an acquisition unit that acquires target learning data from specified time series data, a division unit that divides the learning data acquired by the acquisition unit to construct subspaces based on explanatory variables included in the learning data, and a compression unit that compresses the learning data for each subspace constructed by the division unit to generate compressed learning data to be used for training a regression model. [Effects of the Invention]

[0007] The present invention has the effect of facilitating model learning with high accuracy. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a diagram illustrating an overall image of processing by a data compression device according to this embodiment. [Figure 2] FIG. 2 is a diagram illustrating biased time series data. [Figure 3] FIG. 3 is a diagram showing an example of the configuration of a data compression device according to this embodiment. [Figure 4] FIG. 4 is a table diagram showing an example of time-series data according to this embodiment. [Figure 5] FIG. 5 is a table showing an example of learning data according to this embodiment. [Figure 6] FIG. 6 is a table showing an example of compressed learning data according to this embodiment. [Figure 7] FIG. 7 is a diagram showing an example of data compression processing according to this embodiment. [Figure 8] FIG. 8 is a diagram showing an example of data compression processing according to this embodiment. [Figure 9] FIG. 9 is a flowchart showing an example of the procedure of the data compression process according to this embodiment. [Figure 10]FIG. 10 is a hardware configuration diagram showing an example of a computer that realizes the functions of the data compression device according to this embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0009] Hereinafter, an embodiment (hereinafter referred to as "embodiment") will be described with reference to the drawings. In the following description, common components are given the same reference numerals, and repeated description will be omitted. Furthermore, the description of the embodiment described below does not limit the data compression device, data compression method, and data compression program according to the present invention.

[0010] <Introduction> First, an introduction to this embodiment will be given. Fig. 1 is a diagram illustrating an overview of processing by a data compression device 100 according to this embodiment. The data compression device 100 shown in Fig. 1 is an example of a computer that provides a technology for constructing subspaces for training data used for model training (hereinafter, sometimes simply referred to as "training data") and then compressing data for each subspace.

[0011] (background) A technology is known that uses a regression model that learns the relationships between time-series data collected by sensors installed in chemical plants and the like to continuously determine characteristic values ​​that are difficult to detect or measure as signs of abnormality.

[0012] When detecting signs of anomalies, a predetermined regression model to be used for the detection is learned in advance. A reference technique is known for efficiently constructing a nonlinear regression model to be learned by calculating and evaluating an evaluation value of the nonlinear regression model.

[0013] However, in the reference technology, time series data collected by sensors or the like is used as is when constructing a nonlinear regression model, making it difficult to perform accurate learning when there is a bias in the data distribution.

[0014] Here, an example of biased time series data will be described. Fig. 2 is a diagram for explaining biased time series data. Fig. 2 shows time series data ((1) in Fig. 2) collected by sensors or the like installed in a chemical plant or the like, and a nonlinear regression space ((2) in Fig. 2) constructed using the time series data.

[0015] Chemical plants and the like tend to have long periods of steady operation under a constant load, and short periods of unsteady state during which operations such as switching the load are performed. For this reason, time series data such as that shown in Figure 2 (1-1) may be collected by sensors installed in chemical plants. Of the time series data shown in Figure 2 (1-1), approximately 90% is time series data collected under steady state conditions, and the remaining approximately 10% is time series data collected under unsteady state conditions.

[0016] Here, when constructing a regression model in a chemical plant that predicts the output sensor value (objective variable) shown in (1-2) of Figure 2 from the input sensor value (explanatory variable) shown in (1-1) of Figure 2, the space of the constructed regression model will be a graph with the horizontal axis representing the input and the vertical axis representing the output, as shown in (2) of Figure 2.

[0017] As mentioned above, time-series data collected in a chemical plant may be biased toward the steady-state portion of the steady-state data. Therefore, in the regression model space shown in (2) of Figure 2, the data related to the steady state is biased ((2-1) and (2-2) of Figure 2).

[0018] In general regression model construction methods such as neural networks, each point in the training data is evaluated with the same importance, so if a regression model is constructed using data with a biased distribution as is, it will be affected by the bias. For example, in the time-series data collected in the above-mentioned chemical plant, there is a lot of data relating to the steady state and little data relating to the unsteady state, so training is performed by placing excessive importance on the steady state portion, resulting in problems such as low accuracy in the unsteady state portion.

[0019] Furthermore, as the amount of training data increases, the computational costs, memory usage, and other computer resources required for training increase, making efficient training more difficult.

[0020] (Overall Overview of Processing by Data Compression Device 100) Therefore, the data compression device 100 according to this embodiment constructs a subspace of the space of the regression model consisting of the objective variable and the explanatory variables along the axis of the explanatory variables for the acquired training data, and compresses the training data contained in the subspace.

[0021] 1, a series of processing steps of the data compression device 100 will be described. First, the data compression device 100 acquires target learning data from predetermined time-series data collected by an external sensor or the like and stored in the storage unit 120 ((1) in FIG. 1).

[0022] The data compression device 100 divides the acquired training data based on explanatory variables included in the training data to construct subspaces ((2) in FIG. 1). Next, the data compression device 100 compresses the training data for each constructed subspace to generate compressed training data (hereinafter, sometimes simply referred to as "compressed training data") to be used for training a regression model ((3) in FIG. 1).

[0023] As described above, the data compression device 100 according to this embodiment can compress the amount of data while retaining information for learning the regression model by utilizing the fact that the learning data has a bias in the regression model space. Therefore, the data compression device 100 can generate compressed learning data that allows a regression model to be constructed efficiently (with low computer resources) and with high accuracy. This allows the data compression device 100 to achieve the effect of facilitating accurate model learning.

[0024] <Description of Data Compression Device 100> (Data Compression Device 100) Next, detailed functions of the data compression device 100 according to this embodiment will be described. Fig. 3 is a diagram showing an example of the configuration of the data compression device 100 according to this embodiment. As shown in Fig. 3, the data compression device 100 has a communication unit 110, a storage unit 120, and a control unit 130.

[0025] The data compression device 100 also has an input unit (not shown) such as a keyboard or touch panel for receiving input from a user, etc., and a display unit (not shown) such as a display or printer for displaying the results of information processing by the data compression device 100 to a user, etc.

[0026] (Communication unit 110) The communication unit 110 performs communication related to the output of learning data, compressed learning data, and information related to the learned regression model, and the input of time-series data, etc. The communication unit 110 is realized by a NIC (Network Interface Card) or the like. The communication unit 110 is connected to a network via wire or wirelessly as necessary, and can transmit and receive information bidirectionally.

[0027] (Storage unit 120) The storage unit 120 is realized by, for example, a semiconductor memory element such as a random access memory (RAM) or a flash memory, or a storage device such as a hard disk or an optical disk. The storage unit 120 stores data and programs used for various processes by the control unit 130. As shown in FIG. 3 , the storage unit 120 has a time-series data DB 121, a training data DB 122, and a training compressed data DB 123.

[0028] (Time series data DB121) The time-series data DB 121 is a database that stores time-series data collected by an external data collection device such as a sensor. An example of the time-series data stored in the time-series data DB 121 will now be described with reference to Fig. 4. Fig. 4 is a table diagram showing an example of the time-series data according to this embodiment.

[0029] The time-series data DB 121 stores "No.", which is information for identifying individual data tables included in the time-series data, in association with the names of the acquired time-series data in a table format, etc. For example, as shown in Fig. 4, the time-series data DB 121 stores time-series data A "a" identified by No. "1."

[0030] Note that "a" shown in the item of time series data A identified by No. "1" is data included in the target item, and may be information such as numerical values, text, symbols, etc. Also, as shown in Fig. 4, the time series data DB 121 can store multiple time series data such as time series data A, B, C, etc.

[0031] (Learning data DB122) The learning data DB 122 is a database that stores learning data extracted from time-series data as data that satisfies predetermined conditions. An example of the learning data stored in the learning data DB 122 will now be described with reference to Fig. 5. Fig. 5 is a table diagram showing an example of the learning data according to this embodiment.

[0032] The learning data DB 122 stores "No.", which is information for identifying individual data included in the learning data, in association with explanatory variables and objective variables in a table format, etc. For example, as shown in Fig. 5, the learning data DB 122 stores an explanatory variable (time-series data A) "a" and an objective variable (time-series data B) "d" identified by No. "1."

[0033] Note that the "a" shown in the item of time series data A identified by No. "1" and the "b" shown in the item of time series data B are examples of data contained in the target item, and may be, for example, information such as numbers, text, symbols, etc.

[0034] The training data stored in the training data DB 122 is represented in a graph format with the first axis (vertical axis) representing the "objective variable" and the second axis (horizontal axis) representing the "explanatory variable," as shown in (1) of Fig. 5. If the training data is multivariate including three or more explanatory variables or objective variables, it is represented by a graph according to the number of dimensions.

[0035] (Learning compressed data DB123) The training compressed data DB 123 is a database that stores training compressed data, which is training data compressed by the compression unit 135 described below. An example of compressed data stored in the training compressed data DB 123 will now be described with reference to Fig. 6. Fig. 6 is a table diagram showing an example of training compressed data according to this embodiment.

[0036] The training compressed data DB 123 stores, in a table format or the like, an item of "subspace," which is information for identifying a subspace constructed by the division unit 134 described below, a response variable, and an explanatory variable in association with each other. For example, as shown in Fig. 6, the training compressed data DB 123 stores an explanatory variable (time-series data A) "A" and a response variable (time-series data B) "D," which are identified by a subspace "1."

[0037] Note that the "A" shown in the item of time series data A identified by the subspace "1" and the "B" shown in the item of time series data B are examples of data included in the target items, and may be, for example, information such as numbers, text, symbols, etc.

[0038] The training compressed data stored in the training compressed data DB 123 is expressed in a graph format with the first axis (vertical axis) representing the "objective variable" and the second axis (horizontal axis) representing the "explanatory variable," as shown in (1) of Fig. 6. Furthermore, if the training compressed data is multivariate including two or more explanatory variables, it is expressed in a graph according to the number of dimensions ((2) of Fig. 6).

[0039] (control unit 130) The control unit 130 is realized by a processor, a micro processing unit (MPU), a central processing unit (CPU), or the like executing various programs stored in the storage unit 120 using RAM as a work area. The control unit 130 is also realized by an integrated circuit (IC) such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA). As shown in FIG. 3 , the control unit 130 includes a reception unit 131, an acquisition unit 132, a preprocessing unit 133, a division unit 134, a compression unit 135, a learning unit 136, and an output unit 137.

[0040] (Reception Department 131) The receiving unit 131 receives time-series data input from sensors, field devices, etc. via the above-mentioned input unit (not shown). Then, the receiving unit 131 stores the received time-series data in the storage unit 120 (time-series data DB 121).

[0041] The accepting unit 131 may accept data periodically or automatically according to preset conditions. The above-mentioned "periodically" means a predetermined time interval, such as "every XX minutes, every XX hours, or every XX days." The "preset conditions" include, for example, a condition such as "accept time-series data between 8:00 AM and 8:00 PM every day," and are not limited to any condition used to determine whether or not to accept time-series data.

[0042] (Acquisition part 132) The acquiring unit 132 acquires, as learning data, data that satisfies a predetermined condition from the time-series data stored in the time-series data DB 121. For example, the acquiring unit 132 acquires, as learning data, data to be used as a target objective variable and explanatory variables from the time-series data. Then, the acquiring unit 132 stores the acquired learning data in the learning data DB 122.

[0043] Furthermore, when acquiring learning data, the acquiring unit 132 may exclude some data or combine data relating to multiple periods. For example, the acquiring unit 132 may acquire, as learning data, data for a target period based on a preset acquisition target period or exclusion target period from the time-series data.

[0044] In the present embodiment, the acquiring unit 132 acquires learning data from time-series data collected by sensors installed in a chemical plant or the like, but the data to be acquired is not limited to time-series data. For example, the acquiring unit 132 can acquire learning data from data that can be used as learning data for constructing a regression model, such as property data for each production lot.

[0045] Furthermore, the acquiring unit 132 can acquire learning data not only from sensor data but also from virtual sensor values ​​calculated based on a predetermined model and data on which moving average processing or the like has been performed.

[0046] (Preprocessing unit 133) The preprocessing unit 133 performs predetermined preprocessing on the learning data when preset conditions are satisfied or when a user issues a command to perform preprocessing. Specifically, the preprocessing unit 133 performs preprocessing on the learning data acquired by the acquiring unit 132 using at least one of noise reduction, dimensional compression, variable transformation, and time delay correction.

[0047] For example, the preprocessing unit 133 removes noise from the acquired training data by calculating a moving average, and also compresses the dimensions of the acquired training data using principal component analysis or an autoencoder.

[0048] Furthermore, the pre-processing unit 133 performs variable transformation on the acquired learning data by performing normalization, discretization, etc. Furthermore, the pre-processing unit 133 performs generation of lag variables, etc., to convert the acquired learning data into data that takes time delay into account.

[0049] The above-mentioned predetermined conditions may be any conditions such as whether or not there is noise in the data, whether or not the data size exceeds a specified size, whether or not a time delay occurs, and the like.

[0050] (Divided part 134) The dividing unit 134 divides the training data to construct a subspace based on the explanatory variables included in the training data acquired by the acquiring unit 132. Specifically, the dividing unit 134 equally divides a data interval determined by the minimum and maximum values ​​of the explanatory variables included in the training data or the upper and lower limit values ​​of the device that acquired the training data to construct a subspace.

[0051] For example, the dividing unit 134 can construct a subspace by equally dividing a data interval between the maximum and minimum values ​​of each explanatory variable included in the acquired learning data. Also, the dividing unit 134 can construct a subspace by equally dividing a data interval between the maximum and minimum values ​​in the specifications of a measuring instrument that collects time-series data or the like.

[0052] Furthermore, the dividing unit 134 divides the training data according to the density of the training data to construct subspaces. For example, the dividing unit 134 can construct subspaces such that the number of data points included in each explanatory variable included in the acquired training data is equal for each subspace.

[0053] (Compression section 135) The compression unit 135 compresses the training data for each subspace constructed by the division unit 134 to generate compressed training data to be used for training a regression model. Then, the compression unit 135 stores the generated compressed training data in the training compressed data DB 123.

[0054] The compression unit 135 calculates compressed values ​​of the objective variables based on a predetermined calculation method for the objective variables for each subspace divided by the division unit 134. Next, the compression unit 135 calculates compressed values ​​of the explanatory variables for each subspace in accordance with the predetermined calculation method used to calculate the compressed values ​​of the objective variables, and generates compressed data for training.

[0055] The compression unit 135 may perform compression using a method set for each explanatory variable. Furthermore, when the explanatory variables are multidimensional, the compression unit 135 can compress the values ​​of the objective variables included in the subspaces divided along the axes of each explanatory variable.

[0056] Next, we will further explain the compression method of the objective variable and the explanatory variables. First, we will explain the compression of the objective variable in the following items. Note that the compression unit 135 can calculate the value of the compressed objective variable by using one or a combination of the methods described below.

[0057] For example, the compression unit 135 calculates, as the value of the compressed objective variable, the average value of the objective variable for each subspace divided by the division unit 134. Then, the compression unit 135 outputs the calculated average value as the compressed objective variable.

[0058] Furthermore, the compression unit 135 calculates the median value of the objective variable for each subspace divided by the division unit 134 as the value of the compressed objective variable. Then, the compression unit 135 outputs the calculated median value as the compressed objective variable.

[0059] Furthermore, the compression unit 135 calculates, as the values ​​of the compressed objective variables, the maximum and minimum values ​​of the objective variables for each subspace divided by the division unit 134. Then, the compression unit 135 outputs the calculated maximum and minimum values ​​as the compressed objective variables.

[0060] Furthermore, the compression unit 135 calculates, as the value of the compressed objective variable, a sampled value according to the distribution (for example, a sampled value for each quartile (quartile)) for the objective variable for each subspace divided by the division unit 134. Then, the compression unit 135 outputs the sampled value as the compressed objective variable.

[0061] Next, compression of explanatory variables will be described. For example, the compression unit 135 calculates the median value of the lattice that constitutes the subspace as the value of the compressed explanatory variable for the explanatory variable of each subspace divided by the division unit 134. Then, the compression unit 135 outputs the calculated explanatory variable as the compressed explanatory variable.

[0062] Furthermore, the compression unit 135 calculates, as the value of a compressed explanatory variable, the average value of the explanatory variables included in each subspace divided by the division unit 134. Then, the compression unit 135 outputs the calculated average value of the explanatory variables as the compressed explanatory variable.

[0063] Furthermore, the compression unit 135 calculates a specific value included in the subspace for the explanatory variable for each subspace divided by the division unit 134 as the value of the compressed explanatory variable. Specifically, when the maximum value, minimum value, quantile, or the like of the objective variable included in the subspace is selected, the compression unit 135 determines the original position of the objective variable as the value of the explanatory variable. Then, the compression unit 135 outputs the determined value as the compressed explanatory variable.

[0064] In the above items, the compression unit 135 can select a compression method for the explanatory variables depending on the compression method for the objective variable. For example, when the compression unit 135 calculates the value of the compressed objective variable as the "average value," it calculates the median value or the maximum and minimum values ​​of the lattice that constitutes the subspace as the compressed explanatory variable. Furthermore, when the compression unit 135 calculates the value of the compressed objective variable as the "maximum and minimum value" or "quantile," it calculates it as a specific value (original value) included in the subspace.

[0065] On the other hand, in addition to processing the values ​​contained in the subspaces described above, the compression unit 135 randomly samples data contained in the subspaces divided by the division unit 134, and calculates the obtained sampled values ​​as the values ​​of the compressed objective variables. For example, the compression unit 135 can generate compressed training data by using random sampling to obtain a certain number or a certain percentage of data from the training data.

[0066] (Learning Section 136) The learning unit 136 learns a regression model using the compressed learning data compressed by the compression unit 135. For example, using the generated compressed learning data, the learning unit 136 can learn (train) the regression model so that when an output value (explanatory variable) of a first sensor is input to the regression model, the regression model outputs an output value (objective variable) of a second sensor.

[0067] (output unit 137) The output unit 137 outputs the data and regression model obtained by the processing executed by the data compression device 100 to a user, an external device, etc. For example, the output unit 137 can output at least one of the learning data acquired by the acquisition unit 132, the compressed learning data generated by the compression unit 135, and the trained regression model trained by the training unit 136.

[0068] (Example of processing) The data compression process implemented by the data compression device 100 according to this embodiment will now be described with reference to Fig. 7 and Fig. 8. Fig. 7 and Fig. 8 are diagrams showing an example of the data compression process according to this embodiment.

[0069] (Generating training data for neural networks) First, "generation of training data for neural network" will be explained using Fig. 7. Fig. 7 shows an example in which training data 122a for a one-input, one-output neural network is compressed to generate compressed training data 123a.

[0070] First, the data compression device 100 acquires the learning data 122a based on predetermined conditions. Next, the data compression device 100 equally divides the data interval determined by the minimum and maximum values ​​of the explanatory variables ((1) in FIG. 7).

[0071] The data compression device 100 calculates the average value of the objective variables included in each divided subspace (for example, (2-1) in FIG. 7). Next, the data compression device 100 identifies the center value of the lattice of the subspace based on the compression method of the explanatory variables corresponding to the compression method (average value) of the objective variables (for example, (2-2) in FIG. 7). Then, the data compression device 100 generates compressed training data 123a in which the average value of the objective variables included in the subspace and the center value of the lattice are used as compressed explanatory variables.

[0072] Although an example of one input and one output has been described in FIG. 7, the data compression device 100 can similarly generate compressed training data even when the training data has multiple inputs and multiple outputs.

[0073] (Generating training data for fuzzy neural networks) Next, "Generation of training data for a fuzzy neural network" will be explained using Fig. 8. Fig. 8 shows an example of compressing training data 122b for a one-input, one-output fuzzy neural network to generate compressed training data 123b.

[0074] First, let us assume that in a typical regression model, the estimated value may be determined as a single point, whereas in a fuzzy neural network, a range of estimated values ​​can be obtained according to a given confidence level (see, for example, Reference 1).

[0075] (Reference 1): Hisao Ishibuchi, Koichi Morioka, Hideo Tanaka (1995) "Learning Trapezoidal Fuzzy Number Connection Weights in Fuzzy Neural Networks" Journal of the Japan Society for Fuzzy Theory and Systems, Vol. 7, No. 2, pp. 347-360<URL:https: / / www.jstage.jst.go.jp / article / jfuzzy / 7 / 2 / 7_KJ00002088402 / _pdf / -char / ja> ,<Searched on June 20, 2020>

[0076] In fuzzy neural networks, in addition to explanatory variables and objective variables, reliability values ​​are used as training data. Reliability is a variable that ranges from 0 to 1, with 0 representing the outermost range of estimated values ​​and 1 representing the innermost range of estimated values. For example, in the graph shown in Figure 2(2), the "dotted line" indicates "reliability h = 0.0 (upper limit)" and "reliability h = 0.0 (lower limit)." Also, in the graph shown in Figure 2(2), the "solid line" indicates "reliability h = 1.0 (upper limit)" and "reliability h = 1.0 (lower limit)."

[0077] Next, we will explain the series of processes for generating training data for a fuzzy neural network. First, the data compression device 100 acquires training data 122b based on predetermined conditions. Next, the data compression device 100 equally divides a data interval determined by the minimum and maximum values ​​of the explanatory variables ((1) in FIG. 8).

[0078] The data compression device 100 calculates the minimum value (reliability 0) and the maximum value (reliability 0) of the objective variable included in each divided subspace (for example, (2-1) in FIG. 8). The data compression device 100 also calculates the average value (reliability 1) of the objective variable included in each divided subspace (for example, (2-2) in FIG. 8).

[0079] Next, the data compression device 100 identifies the central value of the lattice of the subspace based on the compression method of the explanatory variables corresponding to the compression method of the objective variable (minimum value (reliability 0), maximum value (reliability 0), average value (reliability 1)).The data compression device 100 then generates compressed training data 123b in which the (minimum value (reliability 0), maximum value (reliability 0), average value (reliability 1)) of the objective variable included in the subspace and the central value of the lattice are used as compressed explanatory variables.

[0080] Although an example of one input and one output has been described in FIG. 8, the data compression device 100 can similarly generate compressed training data even when the training data has multiple inputs and multiple outputs.

[0081] (Processing Procedure) From here, the processing procedure by the data compression device 100 according to this embodiment will be described with reference to Fig. 9. Fig. 9 is a flowchart showing an example of the procedure of the data compression processing according to this embodiment.

[0082] The receiving unit 131 receives time-series data (S101). Next, the acquiring unit 132 acquires target learning data from the time-series data (S102).

[0083] If the learning data is to be preprocessed (Yes in S103), the preprocessing unit 133 executes the preprocessing of the learning data (S104). If the learning data is not to be preprocessed (No in S103), the data compression device 100 skips the step of S104.

[0084] The dividing unit 134 divides the explanatory variables of the training data to construct subspaces (S105). The compressing unit 135 compresses the training data divided for each divided subspace to generate compressed training data (S106).

[0085] Then, the process ends in the data compression device 100. The data compression device 100 can execute a learning process using the generated compressed learning data, and an output process to a user or an external device (not shown in the flowchart).

[0086] (effect) Next, we will explain the effects of the data compression device 100 according to this embodiment. When the data used to construct a general regression model such as a neural network has a biased distribution, the accuracy of learning may decrease. Furthermore, the more learning data there is, the more computer resources are required, making efficient learning more difficult.

[0087] Therefore, the acquisition unit 132 of the data compression device 100 according to this embodiment acquires target training data from predetermined time-series data. The division unit 134 of the data compression device 100 divides the training data based on explanatory variables included in the acquired training data to construct subspaces. The compression unit 135 of the data compression device 100 compresses the training data for each constructed subspace to generate compressed training data to be used for training a regression model.

[0088] As described above, the data compression device 100 compresses the training data in a manner that eliminates bias in the space of the regression model, thereby generating training data that reduces the computer resources required for training and simultaneously improves the accuracy of the regression model.

[0089] For example, by executing the data compression method according to this embodiment, the data compression device 100 achieves the effect of reducing the computer resources required to construct a regression model by compressing portions of the training data that have a biased distribution.

[0090] Furthermore, the data compression device 100 eliminates the bias of data in the regression space, thereby solving the problem of intensive learning of areas with high data density. As a result, the data compression device 100 has the effect of improving the accuracy of the regression model.

[0091] Therefore, the data compression device 100 of this embodiment has the effect of facilitating accurate model learning.

[0092] Specifically, the dividing unit 134 constructs subspaces by equally dividing a data interval determined by the minimum and maximum values ​​of explanatory variables included in the training data or the upper and lower limit values ​​of the device that acquired the training data. The dividing unit 134 also constructs subspaces by dividing the training data according to the density of the training data.

[0093] Through the above-described processing, the data compression device 100 can construct a subspace according to the number of dimensions based on an optimal method for constructing a subspace according to the input time series data. By constructing a subspace according to the number of dimensions, the data compression device 100 can appropriately compress various types of time series data having biased distributions. Therefore, the data compression device 100 can appropriately generate compressed training data according to the input time series data.

[0094] The compression unit 135 calculates compressed values ​​of the explanatory variables for each subspace divided by the division unit 134, based on a predetermined calculation method. For example, the compression unit 135 calculates, as the value of the compressed explanatory variables, at least one of the mean, median, maximum and minimum values, and a sampled value according to the distribution, for the explanatory variables for each subspace divided by the division unit 134. Then, the compression unit 135 calculates compressed values ​​of the explanatory variables for each subspace, based on the predetermined calculation method used to calculate the compressed values ​​of the explanatory variables, and generates compressed training data.

[0095] Through the above-described processing, the data compression device 100 can compress the objective variables using an optimal objective variable compression method according to the type of acquired learning data, the purpose of learning, etc. Therefore, the data compression device 100 can generate data to be used for learning, which allows for more efficient learning than ever before, by appropriately compressing various types of learning data or appropriately compressing learning data according to the purpose of learning.

[0096] The compression unit 135 calculates the compressed explanatory variable values ​​for each subspace according to the predetermined calculation method used to calculate the compressed objective variable values. For example, the compression unit 135 calculates, as the compressed explanatory variable values, at least one of the median of the lattice points constituting the subspace, the average value of the explanatory variables included in the subspace, and a specific value included in the subspace, for the explanatory variables for each subspace divided by the division unit 134. The compression unit 135 also randomly samples data included in the subspaces divided by the division unit 134, and calculates the obtained sampled values ​​as the compressed objective variable values.

[0097] Through the above-described processing, the data compression device 100 can compress the explanatory variables based on the optimal explanatory variable compression method corresponding to the target variable compression method. Therefore, the data compression device 100 can appropriately compress various types of learning data or appropriately compress the learning data according to the learning purpose, thereby generating data to be used for learning that allows learning to be performed more efficiently than ever before.

[0098] The preprocessing unit 133 performs preprocessing on the learning data acquired by the acquiring unit 132 using at least one of noise reduction, dimensional compression, variable transformation, and time delay correction.

[0099] As described above, the data compression device 100 performs preprocessing to efficiently execute compression processing according to input time-series data. As a result, the data compression device 100 performs compression processing using the preprocessed training data, thereby achieving the effect of enabling appropriate generation of compressed training data.

[0100] The learning unit 136 learns the regression model using the compressed learning data compressed by the compression unit 135. Therefore, the data compression device 100 performs learning using the generated compressed learning data, thereby achieving the effect of improving the accuracy and speed of learning the regression model.

[0101] <Modification> The following describes modified examples realized by the data compression device 100 according to this embodiment.

[0102] (Data, etc.) The time-series data, training data, training compressed data, regression models, subspaces, names of functional parts of the data compression device 100, steps, processes, names of steps or processes, etc. used in the description of the above embodiments are merely examples and can be changed as desired.

[0103] Although the storage unit 120 of the data compression device 100 according to this embodiment has been described as storing target data and information in a table format, the present invention is not limited to this. For example, the storage unit 120 can store target data and information in a predetermined file, image, or other format in addition to the table format.

[0104] For example, the time series data DB121 stores "No.", which is information identifying the data acquired in a time series contained in the time series data, and the name of the acquired time series data, but the items to be stored and the information within the items are not limited.

[0105] For example, the learning data DB122 stores "No.", which is information for identifying individual data included in the learning data, in association with the objective variable and the explanatory variable in a table format, but the items to be stored and the information within the items are not limited.

[0106] (application areas, etc.) The data compression device 100 according to this embodiment can be applied to various fields such as chemical plants, chemical processes, power plant facilities, machining processes (presses, NC (Numerically Controlled) machine tools / MC (Machining Centers)), semiconductor manufacturing processes, and water treatment facilities. Note that the above-mentioned application fields are merely examples, and the application fields of the data compression device 100 according to this embodiment are not limited thereto.

[0107] (Regarding regression models) In this embodiment, the regression model may be constructed based on polynomial approximation using the least squares method, a neural network, a fuzzy neural network, Gaussian process regression, a random forest, etc. Note that the above-described regression model construction method is merely an example, and the regression model construction method of the data compression device 100 according to this embodiment is not limited thereto.

[0108] (Flowcharts, etc.) The steps in the flowcharts may be interchanged as long as there is no contradiction, and some steps may not be performed. In addition, conjunctions such as "next," "continue," "further," "at this time," and "on this occasion" used in the explanation of the flowcharts do not limit the order or timing of the execution of the processes in the flowcharts.

[0109] (others) Of the processes described in the above embodiments and variations, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically using known methods. In addition, the information, including the processing procedures, specific names, various data, and parameters shown in the above documents and drawings, can be changed as desired unless otherwise specified. For example, the various information shown in each drawing is not limited to the information shown.

[0110] Furthermore, the components of each device shown in the figure are conceptual functional components and do not necessarily have to be physically configured as shown in the figure. In other words, the specific form of distribution or integration of each device is not limited to that shown in the figure, and all or part of them can be functionally or physically distributed or integrated in any unit depending on various loads, usage conditions, etc.

[0111] The above-described components include those that can be easily imagined by a person skilled in the art, those that are substantially the same, and those that fall within the so-called equivalent range. Furthermore, the above-described embodiments and modifications can be combined as appropriate within the scope of not causing any contradiction in the processing content.

[0112] Furthermore, the above-mentioned "section, module, unit" can be read as "means" or "circuit," etc. For example, a control section can be read as control means or a control circuit.

[0113] Although some of the embodiments have been described in detail above with reference to the drawings, these are merely examples, and it is possible to implement the present embodiments in other forms that have undergone various modifications and improvements based on the knowledge of those skilled in the art, including the aspects described in the Disclosure of the Invention section.

[0114] <Hardware configuration> The data compression device 100 according to this embodiment is realized, for example, by a computer 1000 configured as shown in Fig. 10. Fig. 10 is a hardware configuration diagram showing an example of a computer that realizes the functions of the data compression device 100 according to this embodiment. The computer 1000 has a configuration in which a CPU 1100, a RAM 1200, a ROM 1300, an auxiliary storage device 1400, a communication I / F (interface) 1500, and an input / output I / F (interface) 1600 are connected by a bus 1800.

[0115] The CPU 1100 operates and controls each unit based on a program stored in the ROM 1300 or the auxiliary storage device 1400. The ROM 1300 stores a boot program executed by the CPU 1100 when the computer 1000 starts up, programs that depend on the hardware of the computer 1000, and the like.

[0116] The auxiliary storage device 1400 stores programs executed by the CPU 1100, data used by the programs, etc. The communication I / F 1500 receives data from other devices via a predetermined communication network NW (including closed-area wireless communication in this embodiment) and sends it to the CPU 1100, and transmits data generated by the CPU 1100 to other devices via the predetermined communication network NW. The CPU 1100 controls output devices such as displays and printers, and input / output devices 1700 such as keyboards and mice, via the input / output I / F 1600. The CPU 1100 acquires data from the input / output devices 1700 via the input / output I / F 1600. The CPU 1100 also outputs generated data to the input / output devices 1700 via the input / output I / F 1600.

[0117] For example, when the computer 1000 functions as various devices according to the present embodiment, the CPU 1100 of the computer 1000 executes a program loaded onto the RAM 1200 to realize the functions of the control unit. [Explanation of symbols]

[0118] 100 Data Compression Device 110 Communications Department 120 Storage section 121 Time Series Data DB 122 Learning Data DB 123 Compressed Data DB for Training 130 control section 131 Reception 132 Acquisition Department 133 Pretreatment section 134 Division 135 Compression section 136 Learning Department 137 Output section

Claims

1. an acquisition unit that acquires target learning data from predetermined time series data; a division unit that divides the learning data acquired by the acquisition unit based on explanatory variables included in the learning data to construct a subspace; a compression unit that compresses the learning data for each of the subspaces constructed by the division unit to generate compressed learning data used for learning a regression model; A data compression device comprising:

2. The dividing unit constructing the subspace by equally dividing a data interval determined by the minimum and maximum values ​​of explanatory variables included in the learning data or the upper and lower limit values ​​of the device that acquired the learning data; and, Dividing the training data according to the density of the training data to construct the subspace; Execute at least one of the following:

2. The data compression device according to claim 1.

3. The compression section calculating a value of a compressed objective variable for each of the subspaces divided by the dividing unit based on a predetermined calculation method; calculating compressed explanatory variable values ​​for each of the subspaces according to the predetermined calculation method used to calculate the compressed response variable values; generating the compressed data for learning; 3. The data compression device according to claim 1 or 2.

4. The compression section calculating, as the value of the compressed objective variable, at least one of an average value, a median value, a maximum value, a minimum value, and a sampled value according to a distribution, for the objective variable for each of the subspaces divided by the dividing unit; 4. The data compression device according to claim 3.

5. The compression section calculating, as the value of the compressed explanatory variable, at least one of a median value of the lattice constituting the subspace, an average value of the explanatory variables included in the subspace, and a specific value included in the subspace, for the explanatory variables for each of the subspaces divided by the dividing unit; 4. The data compression device according to claim 3.

6. The compression section randomly sampling data included in the subspace divided by the dividing unit; The obtained sampled value is calculated as the value of the compressed response variable.

3. The data compression device according to claim 1 or 2.

7. The method further includes a pre-processing unit that performs pre-processing on the learning data acquired by the acquisition unit using at least one of noise reduction, dimensional compression, variable transformation, and time delay correction.

3. The data compression device according to claim 1 or 2.

8. further comprising a learning unit that learns the regression model using the compressed data for learning compressed by the compression unit; 3. The data compression device according to claim 1 or 2.

9. A data compression method to be executed by a data compression device, an acquisition step of acquiring learning data from predetermined time series data; a division step of dividing the learning data acquired in the acquisition step based on explanatory variables included in the learning data to construct a subspace; a compression step of compressing the training data for each of the subspaces constructed by the division step to generate compressed training data to be used for training a regression model; 10. A data compression method comprising:

10. An acquisition procedure for acquiring learning data from predetermined time series data; a division step of dividing the training data based on explanatory variables included in the training data acquired by the acquisition step to construct a subspace; a compression step of compressing the training data for each of the subspaces constructed by the division step to generate compressed training data to be used for training a regression model; A data compression program that causes a computer to execute the above steps.

Images