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

Abstract

The present invention addresses the problem of providing a data compression device, a data compression method, and a data compression program product that facilitate model learning with good precision. A data compression device (100) acquires learning data to be a subject from predetermined time-series data. The data compression device (100) divides the learning data on the basis of the interpretation variables included in the acquired learning data, and constructs subspaces. A data compression device (100) compresses the learning data of each constructed subspace, and generates learning compressed data for learning the regression model.

Description

Technical Field

[0001] This invention relates to a data compression device, a data compression method, and a data compression program product. Background Technology

[0002] Regarding machine learning models, including neural networks such as fuzzy neural networks, there are cases where the machine learning model learns the relationships between time-series data collected by sensors installed in chemical plants, etc., and uses the learned model to detect anomalies or continuously calculate characteristic values ​​that are difficult to measure.

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

[0004] [Existing Technical Documents] [Patent Literature] [Patent Document 1] Japanese Patent Application Publication No. 2022-056746 Summary of the Invention [The problem the invention aims to solve] However, in the aforementioned prior art, it is sometimes difficult to achieve good model learning accuracy. For example, a prior art is a method for constructing a nonlinear regression model and using it for anomaly detection. However, in the prior art, data obtained from sensors, etc., is directly used when constructing the nonlinear regression model. Therefore, when using time series data with biased data distribution as training data, it is difficult to achieve good learning accuracy.

[0005] [Technical means to solve the problem] Therefore, in order to solve the aforementioned problem and achieve the objective, the data compression apparatus of the present invention is characterized by comprising: an acquisition unit for acquiring learning data as an object from a predetermined time series data; a segmentation unit for segmenting the learning data to construct subspaces based on the explanatory variables included in the learning data acquired by the acquisition unit; and a compression unit for compressing the learning data in each of the subspaces constructed by the segmentation unit to generate compressed learning data for learning a regression model.

[0006] [The effects of the invention] According to the present invention, the model can be learned easily and with good accuracy. Attached Figure Description

[0007] Figure 1 This diagram illustrates the overall processing using the data compression apparatus of this embodiment.

[0008] Figure 2 It is a graph used to illustrate time series data with bias.

[0009] Figure 3 This is a diagram illustrating an example of the structure of the data compression apparatus of this embodiment.

[0010] Figure 4 This is a table diagram illustrating an example of time-series data in this embodiment.

[0011] Figure 5 This is a table diagram illustrating an example of learning data used in this implementation method.

[0012] Figure 6 This is a table diagram illustrating an example of compressed data used for learning in this embodiment.

[0013] Figure 7 This diagram illustrates an example of the data compression process in this embodiment.

[0014] Figure 8 This diagram illustrates an example of the data compression process in this embodiment.

[0015] Figure 9 This is a flowchart illustrating an example of the data compression processing sequence in this embodiment.

[0016] Figure 10 This is a hardware structure diagram illustrating an example of a computer that implements the data compression device of this embodiment.

[0017] Explanation of icon numbers 100: Data compression device 110: Ministry of Communications 120: Storage Department 121: Time Series Data DB 122: Learning to use data databases 123: Learning to use compressed data databases 130: Control Department 131: Receiving Department 132: Acquisition Department 133: Pre-processing Department 134: Segmentation 135: Compression section 136: Academic Department 137: Output Department Detailed Implementation Here, embodiments (hereinafter referred to as "implementations") will be described with reference to the drawings. Furthermore, in the following description, common structural elements are labeled with the same reference numerals, and repeated descriptions are omitted. Moreover, the description of the embodiments described below does not limit the data compression apparatus, data compression method, and data compression program of the present invention.

[0018] <Preface> First, a preface to this embodiment will be presented. Figure 1 This is a diagram illustrating the overall processing of the data compression apparatus 100 according to this embodiment. Figure 1 The data compression device 100 shown is an example of a computer that provides learning data (hereinafter sometimes simply referred to as "learning data") used for learning a model and compresses the data separately for each subspace after the subspace is constructed.

[0019] (background) There are known techniques that use regression models that learn the relationships between time-series data collected by sensors installed in chemical plants, etc., to detect abnormal signs or continuously calculate characteristic values ​​that are difficult to measure.

[0020] In cases such as detecting abnormal signs, a predetermined regression model used for the detection is learned in advance. Therefore, there are known reference techniques for efficiently constructing a nonlinear regression model of a learning object by calculating an evaluation value of the nonlinear regression model.

[0021] However, in the reference technology, time series data collected by sensors or the like are directly used when constructing the nonlinear regression model, making it difficult to learn with good accuracy when the data distribution is biased.

[0022] Here, we will illustrate an example of time series data with bias. Figure 2 It is a graph used to illustrate time series data with bias. Figure 2 This shows time-series data collected by sensors installed in chemical plants, etc. Figure 2 (1) and the nonlinear regression space constructed using the time series data ( Figure 2 (2)).

[0023] Chemical plants and similar facilities tend to have long steady-state periods operating under a certain load, and short unsteady-state periods during load switching and other operations. Therefore, there is a need to collect data using sensors installed in chemical plants, such as... Figure 2 The situation of time series data shown in (1-1). In Figure 2In the time series data shown in (1-1), about 90% of the data is collected in a steady state, and the remaining 10% is collected in a non-steady state.

[0024] Here, in the chemical plant, in the construction based on Figure 2 The input sensor values ​​(explanatory variables) predicted in (1-1) Figure 2 In the case of a regression model for the sensor values ​​(response variable) output as shown in (1-2), the space of the constructed regression model becomes as follows: Figure 2 The graph shown in (2) is a graph with the horizontal axis representing input and the vertical axis representing output.

[0025] As mentioned above, time-series data collected in chemical plants exhibit a bias towards the steady-state portion, present in both steady-state and non-steady-state conditions. Therefore, in Figure 2 In the space of the regression model shown in (2), there exists a partial set of data related to steady state. Figure 2 (2-1) and (2-2)).

[0026] Typical regression model construction methods, such as those using neural networks, assign equal importance to all points in the learning data for evaluation. Therefore, if biased data is used directly to build the regression model, it will be affected by the bias. For example, in the time series data collected in the chemical plant, there is more data related to steady state and less data related to non-steady state. This leads to an overemphasis on the steady state portion during learning, resulting in problems such as lower accuracy in the non-steady state portion.

[0027] Moreover, the more learning data there is, the more computing resources related to learning, such as computational costs and memory usage, are required, making it more difficult to conduct efficient learning.

[0028] (Overall overview of the processing using the data compression device 100) Therefore, the data compression apparatus 100 of this embodiment compresses the learning data included in the subspace of the regression model, which includes response variables and explanatory variables, along the axis of the explanatory variables.

[0029] Return here Figure 1 The following describes a series of processes handled by the data compression device 100. First, the data compression device 100 acquires learning data (which is the target data) from a predetermined time-series data collected by external sensors and stored in the storage unit 120. Figure 1 (1)).

[0030] The data compression device 100 constructs subspaces by segmenting the acquired learning data based on the explanatory variables included in the learning data. Figure 1(2)). Next, the data compression device 100 compresses the learning data of each constructed subspace to generate compressed learning data (hereinafter sometimes simply referred to as "compressed learning data") for learning the regression model. Figure 1 (3)).

[0031] As described above, the data compression apparatus 100 of this embodiment can retain the information used for learning the regression model while compressing the data volume, even when the learning data has biases within the regression model space. Therefore, the data compression apparatus 100 can generate compressed learning data that can efficiently (with low computer resources) and accurately construct a regression model. Thus, the data compression apparatus 100 achieves the effect of easily and accurately learning the model.

[0032] <Description of Data Compression Device 100> (Data compression device 100) Here, the detailed functions of the data compression device 100 of this embodiment will be described. Figure 3 This is a diagram illustrating an example of the structure of the data compression apparatus 100 according to this embodiment. (See diagram for example.) Figure 3 As shown, the data compression device 100 includes a communication unit 110, a storage unit 120, and a control unit 130.

[0033] Furthermore, the data compression device 100 includes an input unit (not shown) such as a keyboard or touch screen for receiving input from users, or a display unit (not shown) such as a monitor or printer for displaying the results of information processing using the data compression device 100 to users.

[0034] (Communications Department 110) The communication unit 110 performs communications related to the output of learning data, compressed learning data, information related to the learned regression model, and the input of time series data. The communication unit 110 is implemented via a network interface card (NIC). Furthermore, the communication unit 110 can connect to a network via wired or wireless means as needed to transmit and receive information bidirectionally.

[0035] (Storage Department 120) The storage unit 120 is implemented, for example, using semiconductor memory elements such as random access memory (RAM) and flash memory, or storage devices such as hard disks and optical disks. The storage unit 120 stores data and programs used in various processes by the control unit 130. Furthermore, such as Figure 3As shown, the storage unit 120 includes a time series data database (DB) 121, a learning data DB 122, and a learning compressed data DB 123.

[0036] (Time series data DB 121) DB 121 is a database that stores time-series data collected by external data collection machines such as sensors. Here, it is used... Figure 4 An example of time series data stored in time series data DB 121 is illustrated. Figure 4 This is a table diagram illustrating an example of time-series data in this embodiment.

[0037] DB 121 uses the "No" information, which identifies each data table within the time series data, to establish a correspondence with the name of the acquired time series data, and stores it in tabular form. For example, ... Figure 4 As shown, time series data DB 121 stores time series data A "a" identified by No "1".

[0038] Furthermore, the "a" shown in the item of time series data A identified by No. "1" represents the data included in the item of the object, which can be information such as numbers, text, or symbols. Moreover, as... Figure 4 As shown, the time series data DB 121 can store multiple time series data such as time series data A, time series data B, time series data C, etc.

[0039] (Study Data DB 122) DB 122, the learning data database, stores selected learning data from time-series data that meets specified conditions. Here, it uses... Figure 5 An example of learning data stored in learning data DB 122 will be explained. Figure 5 This is a table diagram illustrating an example of learning data used in this implementation method.

[0040] The learning data DB 122 establishes a correspondence between the "No" values, explanatory variables, and response variables used to identify the individual data points included in the learning data, and stores them in tabular form. For example, ... Figure 5 As shown, the learning database DB 122 stores the explanatory variable (time series data A) "a" and the response variable (time series data B) "d" identified by No "1".

[0041] Furthermore, 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 the data included in the item of the object, such as numerical values, text, symbols and other information.

[0042] And, as Figure 5 As shown in (1), the learning data stored in the learning data DB 122 is represented by a chart with the first axis (vertical axis) set as the "response variable" and the second axis (horizontal axis) set as the "explanatory variable". In addition, when the learning data is multivariate and includes three or more explanatory or response variables, it is represented by a chart corresponding to the dimensions.

[0043] (Learning to use compressed data DB 123) The learning compressed data DB 123 is a database that stores learning data compressed by the compression unit 135 described below, i.e., learning compressed data. Here, it is used... Figure 6 An example of compressed data stored in learning compressed data DB 123 is illustrated. Figure 6 This is a table diagram illustrating an example of compressed data used for learning in this embodiment.

[0044] The learning process uses compressed data DB 123 to identify the information in the subspace constructed through the segmentation section 134 described below, namely, the items, response variables, and explanatory variables of the "subspace," establishing a correspondence, and storing it in tabular form, etc. For example, as... Figure 6 As shown, the learning uses compressed data DB 123 to store the explanatory variable (time series data A) "A" and the response variable (time series data B) "D" identified through subspace "1".

[0045] Furthermore, the "A" shown in the item of time series data A identified through subspace "1" and the "B" shown in the item of time series data B are examples of the data included in the item of the object, such as numerical values, text, symbols and other information.

[0046] And, as Figure 6 As shown in (1), the learning compressed data stored in learning compressed data DB 123 is represented in a chart format with the first axis (vertical axis) set as the "response variable" and the second axis (horizontal axis) set as the "explanatory variable". Furthermore, in the case of multivariate learning compressed data including two or more explanatory variables, it is represented in a chart corresponding to the dimensions. Figure 6 (2)).

[0047] (Control Department 130) The control unit 130 is implemented by executing various programs stored in the storage unit 120 using RAM as the operating area, such as a processor, microprocessor (MPU), or central processing unit (CPU). Furthermore, the control unit 130 is implemented, for example, using an integrated circuit (IC) such as an application-specific integrated circuit (ASIC) or a field-programmable gate array (FPGA). Moreover, as... Figure 3 As shown, the control unit 130 includes a receiving unit 131, an acquisition unit 132, a preprocessing unit 133, a segmentation unit 134, a compression unit 135, a learning unit 136, and an output unit 137.

[0048] (Receiving Department 131) The receiving unit 131 receives time-series data input from sensors or field instruments via the 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).

[0049] Furthermore, the receiving unit 131 can also receive data periodically or automatically according to preset conditions. The term "periodically" means, for example, a specified time interval such as "every 00 minutes, every 00 hours, every 00 days". Moreover, the "preset conditions" include, for example, conditions such as "receive time series data between 8:00 AM and 8:00 PM every day", and are not limited as long as they are conditions used to determine whether to perform the reception of time series data.

[0050] (Acquisition Department 132) The acquisition unit 132 acquires data from the time series data stored in the time series data DB 121 that meets specified conditions as learning data. For example, the acquisition unit 132 acquires data from the time series data that are set as response variables and explanatory variables as objects, as learning data. Furthermore, the acquisition unit 132 stores the acquired learning data in the learning data DB 122.

[0051] Furthermore, the acquisition unit 132 can partially exclude data or combine data from multiple periods when acquiring learning data. For example, the acquisition unit 132 can acquire data from time series data that is considered as the target period based on a pre-set acquisition target period or exclusion target period as learning data.

[0052] Furthermore, in this embodiment, the acquisition unit 132 has been described as acquiring learning data from time-series data collected by sensors installed in chemical plants, etc., but the data to be acquired is not limited to time-series data. For example, the acquisition unit 132 can acquire learning data from data such as trait data of each production batch that can be used as learning data for constructing a regression model.

[0053] Furthermore, the acquisition unit 132 can acquire learning data not only from sensor data, but also from virtual sensor values ​​calculated based on a specified model, or from data processed by moving averages, etc.

[0054] (Preprocessing Department 133) The preprocessing unit 133 performs prescribed preprocessing on the learning data when pre-set conditions are met or when instructed by the user to perform preprocessing. Specifically, the preprocessing unit 133 preprocesses the learning data acquired by the acquisition unit 132 using at least one of noise reduction, dimensionality reduction, variable transformation, and time delay correction.

[0055] For example, the preprocessing unit 133 removes noise from the acquired learning data by calculating a moving average. Furthermore, the preprocessing unit 133 uses principal component analysis or an autoencoder to reduce the dimensionality of the acquired learning data.

[0056] Furthermore, the preprocessing unit 133 performs variable transformation on the learning data obtained by performing normalization or discretization. Moreover, the preprocessing unit 133 performs the generation of lag variables, etc., to convert the obtained learning data into data that takes into account time delays.

[0057] Furthermore, as a preset condition, any conditions can be set, such as the presence or absence of data noise, whether the data size exceeds the specified size, or whether a time delay occurs.

[0058] (Divider 134) The segmentation unit 134 constructs a subspace by segmenting the learning data based on the explanatory variables included in the learning data acquired by the acquisition unit 132. Specifically, the segmentation unit 134 constructs a subspace by equally segmenting the data intervals determined by the minimum and maximum values ​​of the explanatory variables included in the learning data, or the upper and lower limits of the machine that acquired the learning data.

[0059] For example, the segmentation unit 134 can equally divide the data interval between the maximum and minimum values ​​of each explanatory variable included in the acquired learning data to construct a subspace. Furthermore, the segmentation unit 134 can equally divide the data interval between the maximum and minimum values ​​of the specifications of the measuring instrument that collects time series data, etc., to construct a subspace.

[0060] Furthermore, the segmentation unit 134 segments the learning data according to the density of the learning data to construct subspaces. For example, the segmentation unit 134 can construct subspaces in such a way that the number of data items included in each explanatory variable of the acquired learning data becomes equal in each subspace.

[0061] (Compression section 135) The compression unit 135 compresses the learning data of each subspace constructed by the segmentation unit 134 to generate compressed learning data for learning the regression model. Furthermore, the compression unit 135 stores the generated compressed learning data in the compressed learning data DB 123.

[0062] The compression unit 135 calculates the value of the compressed response variable for each subspace divided by the segmentation unit 134 based on a predetermined calculation method. Then, the compression unit 135 calculates the value of the compressed explanatory variable for each subspace according to the predetermined calculation method used to calculate the value of the compressed response variable, and generates compressed data for learning.

[0063] Furthermore, the compression unit 135 can perform compression by setting a method for each explanatory variable separately. Moreover, when the explanatory variables are multidimensional, the compression unit 135 can compress the values ​​of the response variables included in the subspace divided along the axes of each explanatory variable.

[0064] Here, the method for compressing the response variable and explanatory variables will be further explained. First, the compression of the response variable will be explained in the following items. In addition, the compression unit 135 can calculate the value of the compressed response variable using one of the methods described below or a combination of multiple methods.

[0065] For example, the compression unit 135 calculates the average value related to the response variables of each subspace divided by the segmentation unit 134 as the value of the compressed response variable. Then, the compression unit 135 outputs the calculated average value as the compressed response variable.

[0066] Furthermore, the compression unit 135 calculates the median value related to the response variables of each subspace divided by the segmentation unit 134 as the value of the compressed response variable. Then, the compression unit 135 outputs the calculated median value as the compressed response variable.

[0067] Furthermore, the compression unit 135 calculates the maximum and minimum values ​​related to the response variables of each subspace divided by the segmentation unit 134 as the values ​​of the compressed response variables. Then, the compression unit 135 outputs the calculated maximum and minimum values ​​as the compressed response variables.

[0068] Furthermore, for the response variable of each subspace divided by the segmentation unit 134, the compression unit 135 calculates the sampled values ​​corresponding to the distribution (e.g., sampled values ​​of each quartile) as the values ​​of the compressed response variable. Then, the compression unit 135 outputs the sampled values ​​as the compressed response variable.

[0069] Next, the compression of explanatory variables will be explained. For example, the compression unit 135 calculates the center value of the grid constituting the subspaces that are related to the explanatory variables of each subspace divided by the segmentation unit 134 as the value of the compressed explanatory variable. Then, the compression unit 135 outputs the calculated explanatory variable as the compressed explanatory variable.

[0070] Furthermore, the compression unit 135 calculates the average value of the explanatory variables included in the subspaces that are related to the explanatory variables of each subspace divided by the segmentation unit 134, and uses this average value as the value of the compressed explanatory variable. Then, the compression unit 135 outputs the calculated average value of the explanatory variables as the compressed explanatory variable.

[0071] Furthermore, the compression unit 135 calculates specific values ​​included in the subspaces related to the explanatory variables of each subspace divided by the segmentation unit 134, and uses these values ​​as the values ​​of the compressed explanatory variables. Specifically, the compression unit 135 determines the original position of the response variable as the value of the explanatory variable after selecting the maximum value, minimum value, quantile, etc., of the response variable included in the subspace. Then, the compression unit 135 outputs the determined value as the compressed explanatory variable.

[0072] In the aforementioned project, the compression unit 135 can select a compression method for the explanatory variables based on the compression method for the response variables. For example, when the value of the compressed response variable is calculated as the "average value," the compression unit 135 calculates the center value, or the maximum and minimum values ​​of the grid constituting the subspace, as the compressed explanatory variables. Furthermore, when the value of the compressed response variable is calculated as the "maximum and minimum values" or "quantiles," the compression unit 135 calculates the specific values ​​(original values) included in the subspace as the values ​​of the compressed explanatory variables.

[0073] On the other hand, in addition to processing the values ​​included in the subspace, the compression unit 135 randomly samples the data included in the subspace divided by the segmentation unit 134, and calculates the obtained sampled values ​​as the values ​​of the compressed response variables. For example, the compression unit 135 uses random sampling to obtain a certain amount or a certain proportion of data from the learning data, thereby generating compressed learning data.

[0074] (Study Department 136) The learning unit 136 uses the compressed learning data compressed by the compression unit 135 to learn a regression model. For example, using the generated compressed learning data, the learning unit 136 can learn (train) the regression model by outputting the output value of the second sensor (response variable) when the output value of the first sensor (explanatory variable) is input into the regression model.

[0075] (Output Section 137) The output unit 137 outputs the data or regression model obtained by the processing performed by the data compression device 100 to the user or an external device. For example, the output unit 137 can output at least one of the learning data acquired by the acquisition unit 132, the learning compressed data generated by the compression unit 135, and the learned regression model learned by the learning unit 136.

[0076] (One example of what was handled) Here, use Figure 7 and Figure 8 The data compression process implemented by the data compression apparatus 100 of this embodiment will be described. Figure 7 and Figure 8 This diagram illustrates an example of the data compression process in this embodiment.

[0077] (Generation of learning data used in neural networks) First, use Figure 7 The generation of learning data for neural networks will be explained. Figure 7 The diagram shows an example of compressing learning data 122a for a single-input, single-output neural network to generate compressed learning data 123a.

[0078] First, the data compression device 100 acquires learning data 122a based on predetermined conditions. Next, the data compression device 100 equally divides the data intervals determined by the minimum and maximum values ​​of the explanatory variables. Figure 7 (1)).

[0079] The data compression device 100 calculates the average value of the response variables included in each of the divided subspaces (e.g., Figure 7 (2-1)). Next, the data compression device 100 determines the center value of the grid of the subspace (e.g., based on the compression method of the explanatory variables corresponding to the compression method (average value) of the response variable. Figure 7 (2-2)). Then, the data compression device 100 generates compressed learning data 123a with the average value of the response variables included in the subspace and the center value of the grid as the compressed explanatory variables.

[0080] In addition, Figure 7The example of single-input single-output is described, but even for multi-input multi-output learning data, the data compression device 100 can generate compressed learning data in the same way.

[0081] (Generation of learning data used in fuzzy neural networks) Next, use Figure 8 The generation of learning data for fuzzy neural networks will be explained. Figure 8 The diagram shows an example of compressing the learning data 122b used in a single-input, single-output fuzzy neural network to generate compressed learning data 123b.

[0082] First, as a premise, in a typical regression model, there exists a point that determines the estimated value. On the other hand, in a fuzzy neural network, the range of the estimated value can be determined based on the provided reliability (see, for example, reference 1).

[0083] (Reference 1): Hisao Ishibuchi, Koichi Morioka, and Hideo Tanaka (1995) "Learning the binding strength of trapezoidal fuzzy numbers in fuzzy neural networks," Journal of the Japanese Fuzzy Society, 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, 2024> In fuzzy neural networks, reliability values ​​are used as training data in addition to explanatory and response variables. Reliability is a variable that takes values ​​from 0 to 1, where 0 represents the outermost range of the estimated value and 1 represents the innermost range. For example, in... Figure 2 In the chart shown in (2), the "dashed lines" represent "reliability h=0.0 (upper limit)" and "reliability h=0.0 (lower limit)". Furthermore, in Figure 2 In the chart shown in (2), the solid line represents "reliability h=1.0 (upper limit)" and "reliability h=1.0 (lower limit)".

[0084] Here, a series of processes for generating the learning data used in the fuzzy neural network are described. First, the data compression device 100 acquires the learning data 122b 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 ( Figure 8 (1)).

[0085] The data compression device 100 calculates the minimum (reliability 0) and maximum (reliability 0) values ​​of the response variables included in each of the divided subspaces (e.g., Figure 8(2-1)). Moreover, the data compression device 100 calculates the average value (reliability 1) of the response variables included in each of the divided subspaces (e.g., Figure 8 (2-2)).

[0086] Next, the data compression device 100 determines the center value of the grid in the subspace based on the compression method of the explanatory variables corresponding to the compression method of the response variables (minimum value (reliability 0), maximum value (reliability 0), average value (reliability 1)). Then, the data compression device 100 generates compressed learning data 123b, using the minimum value (reliability 0), maximum value (reliability 0), average value (reliability 1), and center value of the grid of the response variables included in the subspace as the compressed explanatory variables.

[0087] In addition, Figure 8 The example of single-input single-output is described, but even for multi-input multi-output learning data, the data compression device 100 can generate compressed learning data in the same way.

[0088] (Processing order) Here, use Figure 9 The processing sequence of the data compression device 100 in this embodiment will be explained. Figure 9 This is a flowchart illustrating an example of the data compression processing sequence in this embodiment.

[0089] The receiving unit 131 receives time-series data (S101). Next, the acquiring unit 132 acquires learning data for the target from the time-series data (S102).

[0090] Here, if the preprocessing of the learning data is performed (S103 Yes), the preprocessing unit 133 performs the preprocessing of the learning data (S104). Moreover, if the preprocessing of the learning data is not performed (S103 No), the data compression device 100 skips the step of S104.

[0091] The segmentation unit 134 divides the explanatory variables of the learning data to construct subspaces (S105). The compression unit 135 compresses the learning data segmented according to each of the segmented subspaces to generate compressed learning data (S106).

[0092] Then, the data compression device 100 completes the process. Furthermore, the data compression device 100 can perform learning processing using the generated compressed learning data, or output processing to a user or external device (not shown in the flowchart).

[0093] (Effect) Here, the effects of the data compression device 100 in this embodiment will be explained. When the data used to construct conventional regression models such as neural networks has a distributional bias, the learning accuracy decreases. Furthermore, the more learning data there is, the greater the demand for computer resources, and the more difficult it becomes to perform efficient learning.

[0094] Therefore, in this embodiment, the acquisition unit 132 of the data compression apparatus 100 acquires learning data that is the target from a predetermined time series data. The segmentation unit 134 of the data compression apparatus 100 divides the learning data into subspaces based on the explanatory variables included in the acquired learning data. The compression unit 135 of the data compression apparatus 100 compresses the learning data in each constructed subspace to generate compressed learning data for learning a regression model.

[0095] As described above, the data compression device 100 compresses the learning data in the space of the regression model in a way that eliminates bias, thereby generating learning data that improves the accuracy of the regression model while reducing the computer resources associated with learning.

[0096] For example, by executing the data compression method of this embodiment, the data compression device 100 achieves the following effect: by compressing the biased portion of the learning data, the computer resources required to build the regression model are reduced.

[0097] Furthermore, the data compression device 100 addresses the issue of focusing on the denser parts of the data during learning by eliminating bias in the regression space. Therefore, the data compression device 100 effectively improves the accuracy of the regression model.

[0098] Therefore, the data compression device 100 according to this embodiment can easily and accurately perform model learning.

[0099] Specifically, the segmentation unit 134 constructs a subspace by equally dividing the data interval according to the minimum and maximum values ​​of the explanatory variables included in the learning data, or the upper and lower limits of the machine acquiring the learning data. Furthermore, the segmentation unit 134 constructs a subspace by dividing the learning data according to the density of the learning data.

[0100] Through the aforementioned processing, the data compression device 100 can construct a dimension-corresponding subspace based on the input time-series data and the optimal method for constructing the subspace. Furthermore, by constructing a dimension-corresponding subspace, the data compression device 100 can appropriately compress various time-series data with skewed distributions. Therefore, the data compression device 100 can appropriately generate compressed data for learning based on the input time-series data.

[0101] For the response variables of each subspace divided by segmentation unit 134, compression unit 135 calculates the values ​​of the compressed response variables based on a prescribed calculation method. For example, compression unit 135 calculates at least one of the mean, median, maximum, and minimum values ​​related to the response variables of each subspace divided by segmentation unit 134, and the sampled values ​​corresponding to the distribution, as the values ​​of the compressed response variables. Furthermore, compression unit 135 calculates the values ​​of the compressed explanatory variables for each subspace according to a prescribed calculation method used to calculate the values ​​of the compressed response variables, thereby generating compressed data for learning.

[0102] Through the aforementioned processing, the data compression device 100 can compress the response variable using an optimal compression method corresponding to the type of learning data acquired or the purpose of learning. Therefore, the data compression device 100 can generate learning data that allows for more efficient learning than before by appropriately compressing various types of learning data or by appropriately compressing learning data according to the purpose of learning.

[0103] The compression unit 135 calculates the values ​​of the compressed explanatory variables for each subspace according to a predetermined calculation method used to calculate the values ​​of the compressed response variables. For example, the compression unit 135 calculates at least one of the following as the values ​​of the compressed explanatory variables: the center value of the grid constituting the subspace, the average value of the explanatory variables included in the subspace, and a specific value included in the subspace, which are related to the explanatory variables of each subspace divided by the segmentation unit 134. Furthermore, the compression unit 135 randomly samples data from the subspaces divided by the segmentation unit 134 and calculates the obtained sampled values ​​as the values ​​of the compressed response variables.

[0104] Through this processing, the data compression device 100 can compress the explanatory variables based on the optimal compression method for the explanatory variables corresponding to the compression method for the response variables. Therefore, the data compression device 100 can generate learning data that allows for more efficient learning than ever before by appropriately compressing various types of learning data, or by appropriately compressing learning data according to the learning objectives.

[0105] The preprocessing unit 133 preprocesses the learning data acquired by the acquisition unit 132 using at least one of noise reduction, dimensionality reduction, variable transformation, and time delay correction.

[0106] As described above, the data compression apparatus 100 performs preprocessing on the input time-series data to efficiently perform compression processing. As a result, the data compression apparatus 100 performs compression processing using preprocessed learning data, thereby achieving the effect of appropriately generating compressed learning data.

[0107] The learning unit 136 uses the compressed learning data compressed by the compression unit 135 to learn the regression model. Therefore, by using the generated compressed learning data, the data compression device 100 achieves the effect of improving the learning accuracy and speed of the regression model.

[0108] <Variation Example> Hereinafter, a modified example implemented by the data compression device 100 of this embodiment will be described.

[0109] (Data, etc.) The time series data, learning data, compressed learning data, regression model, subspace, name of functional part of data compression device 100, steps, processes, and names of steps or processes used in the description of the implementation method are ultimately one example and can be arbitrarily changed.

[0110] Furthermore, while the storage unit 120 of the data compression apparatus 100 in this embodiment stores object data or information in tabular form, it is not limited to this. For example, in addition to tabular form, the storage unit 120 can store object data or information in the form of a specified file, image, etc.

[0111] For example, time series data DB 121 stores "No" as information to identify the acquired data and the name of the acquired time series data in the time series included in the time series data, but there are no restrictions on the items stored and the information within the items.

[0112] For example, the learning data DB 122 establishes a correspondence between the "No", response variable and explanatory variable, which are used to identify the individual data included in the learning data, and stores them in tabular form, etc., but there are no restrictions on the items stored and the information within the items.

[0113] (Application areas, etc.) The data compression device 100 of this embodiment can be applied to various fields such as chemical plants, chemical processes, power plant equipment, machining processes (pressing, numerically controlled (NC) machine tools / machining centers and other semiconductor manufacturing processes), and water treatment equipment. However, the application areas described are ultimately just examples, and therefore the application areas of the data compression device 100 of this embodiment are not limited.

[0114] (Regarding regression models) In this embodiment, the regression model can be constructed based on polynomial approximation using the least squares method, neural networks, fuzzy neural networks, Gaussian process regression, random forests, etc. Furthermore, the method for constructing the regression model is ultimately just one example, and therefore the method for constructing the regression model in the data compression device 100 of this embodiment is not limited.

[0115] (Flowcharts, etc.) The steps in a flowchart can be substituted for each other within the bounds of not contradicting each other, and there may also be steps that are not implemented. Moreover, the connecting words such as "next," "next," "furthermore," "at this time," and "at this point" in the flowchart description do not limit the order or timing of the implementation of the processes in the flowchart.

[0116] (other) All or part of the processes described in the embodiments and variations as automatically performed processes can also be performed manually, or all or part of the processes described as manually performed processes can be performed automatically by known methods. Furthermore, information regarding the processing order, specific names, various data, or parameters shown in the documents or figures can be arbitrarily changed, unless specifically stated otherwise. For example, the various information shown in the figures is not limited to the information illustrated.

[0117] Furthermore, the structural elements of each device shown in the illustrations are functional conceptual elements and may not be physically constructed as depicted. That is, the specific form of the distribution or integration of each device is not limited to that shown in the illustrations. Depending on various loads or usage conditions, all or part of them can be functionally or physically distributed or integrated into any unit.

[0118] The structural elements include structural elements that can be easily assumed by those skilled in the art, substantially identical structural elements, and structural elements of equal scope. Furthermore, the embodiments and variations described above can be appropriately combined within the scope where the processing content does not contradict each other.

[0119] Furthermore, the term "section, module, unit" mentioned above can be replaced with "mechanism" or "circuit," etc. For example, the control section can be replaced with a control mechanism or control circuit.

[0120] The above describes several embodiments in detail based on the drawings. However, these embodiments are illustrative and can be represented by the forms described in the disclosure section of the invention. Other forms can be implemented by applying various modifications and improvements based on the knowledge of those skilled in the art.

[0121] <Hardware Structure> The data compression device 100 in this embodiment, for example, uses, as... Figure 10The structure shown is implemented by computer 1000. Figure 10 This is a hardware structure diagram illustrating an example of a computer that implements the functions of the data compression apparatus 100 of this embodiment. The computer 1000 has a configuration in which a CPU 1100, RAM 1200, read-only memory (ROM) 1300, auxiliary storage device 1400, communication I / F (interface) 1500, and input / output I / F (interface) 1600 are connected via a bus 1800.

[0122] The CPU 1100 operates based on the program stored in the ROM 1300 or the auxiliary storage device 1400, and controls the various parts. The ROM 1300 stores the startup program executed by the CPU 1100 when the computer 1000 starts up, or programs that depend on the hardware of the computer 1000.

[0123] Auxiliary storage device 1400 stores programs executed by CPU 1100 and data used by said programs. Communication I / F 1500 receives data from other machines and sends it to CPU 1100 via a defined communication network NW (in this embodiment, this includes wireless communication within a closed area), and sends data generated by CPU 1100 to other machines via the defined communication network NW. CPU 1100 controls output devices such as displays or printers, and input / output devices 1700 such as keyboards or mice via input / output I / F 1600. CPU 1100 acquires data from input / output devices 1700 via input / output I / F 1600. Furthermore, CPU 1100 outputs generated data to input / output devices 1700 via input / output I / F 1600.

[0124] For example, when the computer 1000 functions as a device in this embodiment, the CPU 1100 of the computer 1000 performs the functions of the control unit by executing a program loaded on the RAM 1200.

Claims

1. A data compression device, characterized in that... include: The acquisition department extracts data from the specified time-series data for use in learning the target. The segmentation unit constructs a subspace by segmenting the learning data based on the explanatory variables included in the learning data acquired by the acquisition unit. and The compression unit compresses the learning data in each of the subspaces constructed by the segmentation unit to generate compressed learning data for learning the regression model.

2. The data compression device according to claim 1, characterized in that, The segmented portion performs at least one of the following processes: The subspace is constructed by equally dividing the data intervals determined by the minimum and maximum values ​​of the explanatory variables included in the learning data, or the upper and lower limits of the machine that acquires the learning data; and The learning data is divided into subspaces based on the density of the learning data.

3. The data compression device according to claim 1 or 2, characterized in that: The compression section For the response variables of each subspace divided by the segmentation, the values ​​of the compressed response variables are calculated based on a prescribed calculation method. The values ​​of the compressed explanatory variables for each of the subspaces are calculated according to the prescribed calculation method used to calculate the values ​​of the compressed response variables. Generate the compressed data used for the learning process.

4. The data compression device according to claim 3, characterized in that: The compression section Calculate at least one of the mean, median, maximum, and minimum values ​​associated with the response variable of each of the subspaces to which the segmentation is divided, and the sampled values ​​corresponding to the distribution, as the value of the compressed response variable.

5. The data compression device according to claim 3, characterized in that: The compression section The value of the compressed explanatory variable is calculated as at least one of the following: the center value of the grid constituting the subspace, the average value of the explanatory variables included in the subspace, and a specific value included in the subspace, which are related to the explanatory variables of each of the subspaces divided by the segmentation.

6. The data compression apparatus according to claim 1 or 2, characterized in that: The compression section Random sampling is performed from the data included in the subspace that has been divided by the segmentation. The obtained sampled values ​​are calculated as the values ​​of the compressed response variable.

7. The data compression apparatus according to claim 1 or 2, characterized in that... Also includes: The preprocessing unit preprocesses the learning data acquired by the acquisition unit using at least one of noise reduction, dimensionality reduction, variable transformation, and time delay correction.

8. The data compression apparatus according to claim 1 or 2, characterized in that... Also includes: The learning unit uses the compressed learning data, which has been compressed by the compression unit, to learn the regression model.

9. A data compression method, characterized in that it is a data compression method executed by a data compression device. include: The acquisition process involves obtaining learning data from the specified time-series data. The segmentation process, based on the explanatory variables included in the learning data acquired through the acquisition process, segments the learning data to construct a subspace; and The compression process compresses the learning data in each of the subspaces constructed through the segmentation process to generate compressed learning data for learning the regression model.

10. A data compression program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, the computer executes them in the following order: The acquisition sequence involves retrieving learning data from a specified time-series dataset. The segmentation order is based on the explanatory variables included in the learning data obtained through the acquisition order, and the learning data is segmented to construct a subspace; and The compression sequence compresses the learning data in each of the subspaces constructed by the segmentation sequence to generate compressed learning data for learning the regression model.

Images