Analytical equipment, analytical methods, and analytical programs

Abstract

To facilitate analyzing a complicated data causal relationship.SOLUTION: An extraction unit 131 extracts a collection of neighbor data of specified data from data pieces each having a plurality of feature amounts. The extraction unit 131 extracts k neighbor data by JIT method. A search unit 134 applies a technique of a cause and effect search technique to the collection of neighbor data to search for a cause and effect between the feature amounts. The search unit 134 applies the LiMGAM algorithm to the extracted data to obtain a linear equation representing a cause and effect relation between the feature amounts.SELECTED DRAWING: Figure 4

Description

[Technical Field]

[0001] This invention relates to an analytical device, an analytical method, and an analytical program. [Background technology]

[0002] The demand for time-series data analysis in the manufacturing industry is high, and its applications are diverse, ranging from quality forecasting and equipment anomaly detection to fault detection and operational automation.

[0003] In practical applications, there is a high demand for Explainable Artificial Intelligence (XAI), which not only improves the average accuracy of machine learning models in prediction and anomaly detection, but also provides the rationale behind their decisions and identifies the structure of the system. XAI is particularly important in plants and similar environments, considering the economic and safety risks associated with accidents.

[0004] Within XAI, there is a field called causal search, which involves determining the system structure and causal relationships between features from data.

[0005] Generally, when a causal relationship is unknown beforehand, it can be estimated through controlled experiments such as randomized controlled trials (RCTs) in which only specific variables are manipulated. However, in reality, conducting such experiments is often difficult.

[0006] Statistical causal search is a framework for identifying causal relationships solely from data when experiments are difficult. The application of causal search to time-series data has also been actively researched in recent years (see, for example, Non-Patent Document 1).

[0007] Within the framework of statistical causal inference, by utilizing causal relationships, we can understand how changing a variable that is independent of other variables (intervention) affects the overall data (intervention effect estimation) (see, for example, Non-Patent Literature 2).

[0008] These technologies make it possible to optimize material input amounts in a plant by working backward from the optimal timeframe, and also to reflect the causal relationships discovered in subsequent decision-making, such as management decisions.

[0009] On the other hand, the data that appears in real-world applications such as time series analysis often does not conform to the ideal conditions assumed in conventional time series analysis and causal exploration methods.

[0010] For example, in time series analysis, stationarity is assumed in many models. This means that the time series data being analyzed does not contain time-dependent trend and seasonal components, and that the system's dynamics and state are constant regardless of time. However, actual time series data contains seasonal shifts in sensor values ​​due to temperature changes, discrete state changes in operating conditions such as changes in production volume, and trend components due to the aging of the plant.

[0011] Just-In-Time models have been proposed as a technique to address prediction problems involving such time-series data or complex nonlinear data (see, for example, Non-Patent Document 8). According to Just-In-Time models, by continuously learning from historical data that is close to the current state, it is possible to track changes in distribution and complex nonlinearities.

[0012] Furthermore, several methods are known for causal exploration, one of the most representative being LiNGAM (Linear Non-Gaussian Acyclic Model) (see, for example, Non-Patent Documents 3 and 4). In LiNGAM, a linear model is assumed for the structural equation that represents the causal relationship. On the other hand, methods for nonlinear causal exploration are also known (see, for example, Non-Patent Documents 5, 6, and 7). [Prior art documents] [Non-patent literature]

[0013] [Non-Patent Document 1] Hyvarinen, A., Zhang, K., Shimizu, S., and Hoyer, PO: Estimation of a structural vector autoregression model using non-gaussianity., Journal of Machine Learning Research, Vol. 11, No. 5 (2010) [Non-Patent Document 2] Pearl, J.: Statistical Causal Inference (2009) [Non-Patent Document 3] Shimizu, S., Hoyer, PO, Hyvarinen, A., Kerminen, A., and Jordan, M.: A linear non-Gaussian acyclic model for causal discovery., Journal of Machine Learning Research, Vol. 7, No. 10 (2006) [Non-Patent Document 4] Shimizu, S., Inazumi, T., Sogawa, Y., Hyvarinen, A., Kawahara, Y., Washio, T., Hoyer, PO, and Bollen, K.: DirectLiNGAM: A direct method for learning a linear non-Gaussian structural equation model, The Journal of Machine Learning Research, Vol. 12, pp. 1225-1248 (2011) [Non-Patent Document 5] Hoyer, P., Janzing, D., Mooij, JM, Peters, J., and Scholkopf, B.: Nonlinear causal discovery with additive noise models, Advances in neural information processing systems, Vol. 21, pp. 689-696 (2008) [Non-Patent Document 6] Peters, Jonas and Mooij, Joris M and Janzing, Dominik and Scholkopf, Bernhard: Causal discovery with continuous additive noise models, Journal of Machine Learning Research, Vol. 15, pp. 2009-2053 (2014) [Non-Patent Document 7] Zheng, X., Aragam, B., Ravikumar, PK, and Xing, EP: Dags with no tears: Continuous optimization for structural learning, Advances in Neural Information Processing Systems, Vol. 31, (2018) [Non-Patent Document 8] Shigeru Yamamoto: Just-In-Time Predictive Control: Predictive Control Based on Accumulated Data, Measurement and Control, Vol. 52, No. 10, pp. 878-884 (2013) [Overview of the project] [Problems that the invention aims to solve]

[0014] However, conventional technologies have the problem that they cannot easily analyze the causal relationships of complex data.

[0015] This is because, as mentioned above, in data that appears in practical applications such as time series, the ideal conditions assumed by conventional causal search methods are often violated.

[0016] For example, many physical phenomena that actually occur in plants and the like are nonlinear, and conventional methods that assume linearity in data can only capture causal relationships as a limited approximation.

[0017] On the other hand, nonlinear causal search methods are also known, but the disadvantages of these nonlinear causal search methods include that causal relationships represented by nonlinear functions are not always easy to interpret, the problem of computational cost, and that effectiveness for non-stationary data is not guaranteed in time series extension. [Means for Solving the Problems]

[0018] In order to solve the above-mentioned problems and achieve the object, the analysis device includes: an extraction unit that extracts a set of neighboring data of specified data from data each having a plurality of feature quantities; and a search unit that applies a causal search method to the set of neighboring data to search for causality between the feature quantities. [Effects of the Invention]

[0019] According to the present invention, causal relationships in complex data can be easily analyzed. [Brief Description of Drawings]

[0020] [Figure 1] Figure 1 is a diagram explaining an outline of the analysis device. [Figure 2] Figure 2 is a diagram explaining a decrease in accuracy of a prediction model over time. [Figure 3] Figure 3 is a diagram explaining a causal search method by JIT-LiNGAM. [Figure 4] Figure 4 is a diagram showing a configuration example of the analysis device. [Figure 5] Figure 5 is a diagram showing an example of a history DB. [Figure 6]Figure 6 illustrates the relationship between the target to be predicted and the training data. [Figure 7] Figure 7 shows an example of a causal relationship between features. [Figure 8] Figure 8 illustrates the distribution shift. [Figure 9] Figure 9 shows an example of a JIT-LiNGAM algorithm. [Figure 10] Figure 10 shows an example of a screen displaying the analysis results. [Figure 11] Figure 11 is a flowchart showing the training process flow. [Figure 12] Figure 12 is a flowchart showing the flow of the prediction process. [Figure 13] Figure 13 is a flowchart showing the flow of the causal search process. [Figure 14] Figure 14 shows an example of a computer configuration for running an analysis program. [Modes for carrying out the invention]

[0021] Embodiments of the analytical apparatus, analytical method, and analytical program according to the present invention will be described in detail below with reference to the drawings. However, the present invention is not limited to the embodiments described below.

[0022] [First Embodiment] [Configuration of the first embodiment] As shown in Figure 1, the analytical instrument 10 is included in a system for managing and controlling the production process of products in the plant. The plant includes a chemical plant for producing chemical products.

[0023] As shown in Figure 1, the analysis device 10 receives sensor data and other information from the plant 20a. Based on the received data, the analysis device 10 trains a prediction model to predict recommended operating values, predicts recommended operating values ​​using the trained prediction model, and explores causal relationships between features related to the prediction model (causal analysis).

[0024] The analysis device 10 performs imitation learning by training a predictive model with the operations actually performed by the operator in plant 20a.

[0025] For example, a predictive model learns the amount of raw materials that operators have previously input in a specific process. The predictive model then outputs the recommended amount of raw materials to input. By setting the amount of raw materials according to the output of the predictive model, the operator can mimic past operations.

[0026] Predictive models that utilize this kind of imitation learning can not only support plant operations but also facilitate the transfer of operator skills, optimize operations, and ultimately enable complete automation of operations.

[0027] Furthermore, causal analysis results are crucial for realizing XAI (Explainable AI). Predictive models are a type of AI and are black boxes, making it difficult to intuitively understand from the outside how each feature is handled.

[0028] For example, in a plant, having users (plant operators or management personnel, etc.) confirm the causal relationships between the features used in the predictive model can reduce the risk of equipment damage and alleviate anxiety.

[0029] The analysis device 10 outputs recommended operating values ​​and causal analysis results to the terminal device 21 used by the user. The user (especially the operator) operates the plant 20b according to the recommended operating values. The user can also check the results of the causal analysis via the terminal device 21.

[0030] The analysis device 10 is an information processing device such as a server or a personal computer. The terminal device 21 is an information processing device such as a personal computer, a tablet terminal or a smartphone.

[0031] Plant 20b may be the same as plant 20a, or it may be a different plant with a similar configuration and function.

[0032] Here, we will explain the importance of causal analysis. As mentioned earlier, it is sometimes desirable to estimate causal relationships through experiments such as RCTs, but in reality, it is often difficult to conduct such experiments. Therefore, it is necessary to estimate causal relationships from data alone. Causal relationships between data can be represented by DAG (Directed Acyclic Graph), SCM (Structural Causal Model), etc.

[0033] Furthermore, correlation analysis, which estimates the correlation between data (correlation coefficient, regression coefficient), is easier to perform than causal analysis. On the other hand, correlation analysis has problems such as the occurrence of Simpson's paradox and the difficulty in identifying spurious correlations.

[0034] As shown in Figure 2, the nature of the data changes over time, which can lead to a decrease in the prediction accuracy of the prediction model. Figure 2 illustrates the decrease in the accuracy of the prediction model over time.

[0035] For example, suppose the target production volume of a certain product in the winter of a certain year was 500 tons. And at that point, suppose the analytical device 10 has trained its predictive model to learn the operator's actions.

[0036] Let's assume that six months later, the season has changed to summer, and the target production volume of the product has changed to 100 tons. In this case, because the conditions, including the season, have changed from six months prior, the recommended operating values ​​output by the trained predictive model may deviate from the optimal operating values. In addition to the season and target production volume shown in Figure 2, deterioration of the plant itself may also be a factor in the deviation.

[0037] One way to address the decrease in accuracy of prediction models over time is to use a JIT model (see, for example, Non-Patent Document 8).

[0038] In this embodiment, the JIT model method (hereinafter referred to as the JIT method) is applied to causal analysis. This makes it possible to easily analyze local causal relationships even for time series data where the causal relationships are non-stationary and nonlinear.

[0039] The analysis device 10 uses a method that applies the JIT method to LiNGAM, a causal discovery technique. This method is referred to here as JIT-LiNGAM. Figure 3 is a diagram illustrating the causal discovery method using JIT-LiNGAM.

[0040] As shown in Figure 3, the analysis device 10 extracts neighboring data 52 of the input data 51 to be analyzed using the JIT method. Then, the analysis device 10 assumes that the causal relationships between the features of the extracted neighboring data 52 are linear and performs causal search (LiNGAM).

[0041] The analytical apparatus 10 will be described in detail using Figure 4. Figure 4 is a diagram showing an example of the configuration of the analytical apparatus.

[0042] The analysis device 10 performs training of a prediction model, prediction processing using the prediction model, and causal search processing to explore the causal relationships between features related to the prediction model.

[0043] As shown in Figure 4, the analysis device 10 includes a communication unit 11, a storage unit 12, and a control unit 13.

[0044] The communication unit 11 performs data communication with other devices via the network. For example, the communication unit 11 is a NIC (Network Interface Card).

[0045] The storage unit 12 is a storage device such as an HDD (Hard Disk Drive), SSD (Solid State Drive), or optical disc. Alternatively, the storage unit 12 may be a rewritable semiconductor memory such as RAM (Random Access Memory), flash memory, or NVSRAM (Non-Volatile Static Random Access Memory).

[0046] The memory unit 12 stores the OS (Operating System) and various programs executed by the analysis device 10. The memory unit 12 also stores prediction model information 121 and history DB 122.

[0047] Predictive model information 121 is information such as parameters for constructing a predictive model. For example, if the model is a neural network, predictive model information 121 is the weights and biases of each layer.

[0048] History DB 122 is a database that stores data acquired from plant 20a along with timestamps. Figure 5 shows an example of the history DB. Alternatively, history DB 122 may be a data set obtained by extracting a portion of all data according to a time window.

[0049] As shown in Figure 5, the history DB122 includes time, temperature, pressure, CO2 concentration, and set value.

[0050] Temperature, pressure, and CO2 concentration correspond to the explanatory variables in the prediction model. The set values ​​correspond to the dependent variable in the prediction model. Note that the features targeted for causal exploration may include either explanatory or dependent variables.

[0051] Furthermore, for example, temperature, pressure, and CO2 concentration are sensor values ​​from sensors installed at various locations in plant 20a. The time is a timestamp indicating the time when the temperature, pressure, and CO2 concentration were acquired.

[0052] The set value is a value that is set by the operator's actions. The set value may also be a normalized version of the actually set value. The set value is a value that identifies the operation of the equipment. For example, the set value corresponds to the amount of raw material input.

[0053] For example, Figure 5 shows that at the time "2021 / 11 / 5 13:30:01", the temperature is "40°C", the pressure is "501 hPa", the CO2 concentration is "700 ppm", and the set value is "0.2".

[0054] The control unit 13 controls the entire analysis device 10. The control unit 13 is, for example, an electronic circuit such as a CPU (Central Processing Unit), MPU (Micro Processing Unit), or GPU (Graphics Processing Unit), or an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or FPGA (Field Programmable Gate Array).

[0055] Furthermore, the control unit 13 has internal memory for storing programs and control data that define various processing procedures, and executes each process using the internal memory. In addition, the control unit 13 functions as various processing units when various programs are run. For example, the control unit 13 has an extraction unit 131, a prediction unit 132, an update unit 133, a search unit 134, and a display control unit 135.

[0056] The extraction unit 131 extracts a set of data from the history DB 122 that is close to the input data. The history DB 122 is an example of data where each data has multiple features.

[0057] The input data is pre-specified data to be analyzed and corresponds to one or more records in the history database 122. For example, the input data may be a record corresponding to the current time, or it may be multiple records pointed out to understand the overall causal relationship.

[0058] Furthermore, each record in the history DB122 can be considered as a multidimensional vector whose elements are features. The extraction unit 131 extracts neighboring data based on an index that measures the proximity between records (for example, Euclidean distance, Mahalanobis distance, or cosine similarity).

[0059] The prediction unit 132 calculates the target variable by inputting explanatory variables into the prediction model constructed from the prediction model information 121.

[0060] For example, the prediction unit 132 inputs the temperature, pressure, and CO2 concentration from the history DB 122 into the prediction model and outputs the set values ​​(prediction process).

[0061] The update unit 133 updates the parameters of the prediction model.

[0062] For example, the update unit 133 updates the parameters of the prediction model, i.e., the prediction model information 121, so that the difference between the setting value output by the prediction model and the setting value actually set by the operator is minimized (training process).

[0063] Figure 6 illustrates the relationship between the prediction target and the training data. The training data, which is data up to time t-1, consists of records in the history DB 122 with registered settings. For example, the update unit 133 updates the prediction model using the training data.

[0064] The data at time t is the data to be predicted, and its setting value is not registered. The prediction unit 132 predicts the setting value of the data to be predicted during the prediction process.

[0065] The search unit 134 uses the neighboring data of the input data extracted by the extraction unit 131 to search for causality between features. The search unit 134 applies a causal search method to the set of neighboring data to search for causality between features. For example, the search unit 134 outputs the results of the causal search as a DAG.

[0066] As shown in Figure 6, the causal relationships between features can be very complex in the overall dataset. Figure 7 shows an example of a causal relationship between features.

[0067] In the following explanation, the features will be represented as x1, x2, x3, and x4. In the example in Figure 7, the causal relationship between x1 and x2 is expressed as x2 = -sin(2x1) + e2 (where e2 is noise). Thus, in the example in Figure 7, the causal relationships between the features are all nonlinear.

[0068] Furthermore, causal relationships between features can sometimes cause distribution shifts. Figure 8 illustrates this distribution shift.

[0069] The horizontal and vertical axes in Figure 8 correspond to different features, respectively. Figure 8 clearly shows two distinct distributions. Thus, the causal relationships between features cannot always be represented by a single distribution.

[0070] In response, the search unit 134 approximates the causal relationships based on the nonlinear structural equations with a linear model using neighboring data.

[0071] First, let's consider a nonlinear structural equation model called ANM (Additive Noise Model), which restricts noise to an additive type. In ANM, the feature x i It can be expressed as shown in equation (1), where P is the number of features.

[0072]

number

[0073] The vector representing the features in equation (1) can be expressed as shown in equation (2).

[0074]

number

[0075] However, here e iAssume that each i is an independent, non-Gaussian noise with a mean of zero. The required point x corresponds to the input data. (q) The neighboring region B(x(q),ε)={x|d(x,x (q) In )<ε}, equation (3) is derived from the Taylor expansion. J(·) is the Jacobian matrix.

[0076]

number

[0077] Required point x (q) Taking the expectation of both sides of equation (3) within the section distribution by the neighboring region B(x(q),ε), we obtain equation (4). Here, x and e, which are indicated by bars directly above them, are the expectation values ​​of x and e, respectively. Note that the required point x (q) The neighboring region B(x(q),ε) is extracted by the extraction unit 131.

[0078]

number

[0079] Subtracting both sides of equation (3) and equation (4) yields equation (5).

[0080]

number

[0081] Equation (5) can be considered as a linear structural equation model in LiNGAM. The search unit 134 solves equation (5) by following the LiMGAM algorithm and obtains the matrix J(x (q) )

[0082] Note that x~ (the space directly above x) and e~ (the space directly above e) in equation (5) can be obtained by sample approximation through data centering.

[0083] Based on the foregoing description, the processing of the extraction unit 131 and the search unit 134 in the present embodiment is represented by the algorithm of FIG. 9. FIG. 9 is a diagram illustrating an example of the algorithm of JIT-LiNGAM.

[0084] As shown in FIG. 9, the algorithm includes accumulated data D{x (t) |t=1,…,T-1}, requested point x (q) =x (T) , distance function d(·,·), and number of neighbors K as inputs.

[0085] The accumulated data D is data accumulated in the history DB 122. The requested point x (q) is data specified as input data, and corresponds to one record in the history DB 122. The distance function d is a function for calculating Euclidean distance between vectors, Mahalanobis distance, or cosine similarity. The number of neighbors K is the number of data items included in neighborhood data. Note that the distance function d is not limited to the above example, and any function capable of calculating a distance between data may be used.

[0086] The algorithm outputs a matrix J(x (T) ). According to the matrix J(x (T) ), a DAG can be constructed.

[0087] The extraction unit 131 extracts, from the accumulated data D, a set of K pieces of data (neighborhood set) in ascending order of distance d(x (q) ,x (t) ,x (q) ) to the requested point x (Step 1).

[0088] The search unit 134 performs centering on the neighborhood set to obtain a centered set (Step 2) (processing corresponding to equations (3), (4) and (5)).

[0089] Further, the search unit 134 trains LiNGAM with the data included in the set obtained in Step 2. That is, the search unit 134 obtains a solution of the linear structural equation (elements of the matrix J(x (T) )).

[0090] This allows the search unit 134 to find a linear equation that represents the relationship between the features. For example, the search unit 134 represents each feature as the sum of a value obtained by multiplying at least one of the features by a coefficient and noise.

[0091] The display control unit 135 outputs the results of the search performed by the search unit 134 as the causal analysis results. For example, the display control unit 135 outputs a screen like the one shown in Figure 10 to the terminal device 21. Figure 10 is a diagram showing an example of a screen that displays the analysis results. In this way, the display control unit 135 outputs the search results performed by the search unit 134 as a directed graph.

[0092] As shown in Figure 10, the display control unit 135 displays a list of features on the causal analysis results display screen. In the example in Figure 10, "x0: set value, x1: temperature, x2: pressure, x3: CO2 concentration" are displayed.

[0093] Furthermore, the display control unit 135 displays a graph (DAG) representing the causal relationship between features and a linear equation representing the causal relationship between each feature on the causal analysis result display screen. In the example in Figure 10, the causal relationship between x2 and other features is shown as x2 = 0.8x0 + 1.5x1 + e2 (where e2 is noise).

[0094] Furthermore, the display control unit 135 displays a message on the causal analysis result display screen that corresponds to the results of the causal analysis. For example, the display control unit 135 outputs the search results from the search unit 134 in natural language text. The display control unit 135 is an example of an output unit.

[0095] In the example in Figure 10, a message is displayed stating that "there is a strong positive causal relationship between temperature and pressure." This makes it possible to explain the results of causal analysis to the user in an intuitive and simple manner.

[0096] Furthermore, the results of causal analysis can be used to control the plant in a way that establishes appropriate causal relationships. The results of causal analysis can also be used for variable selection, specifically for choosing the features to input into the predictive model.

[0097] [Processing flow of the first embodiment] Figure 11 illustrates the training process flow using the analysis device 10. Figure 11 is a flowchart showing the training process flow.

[0098] As shown in Figure 11, first, the analyzer 10 extracts data from the history DB 122 that is close to the input data (step S101).

[0099] Then, the analysis device 10 trains the model using the extracted data (step S102). Alternatively, the analysis device 10 may train the model using all the data in the history database 122.

[0100] Figure 12 illustrates the flow of the prediction process performed by the analysis device 10. Figure 12 is a flowchart showing the flow of the prediction process.

[0101] As shown in Figure 12, first, the analyzer 10 inputs the input data into the model and calculates the output data (step S201).

[0102] Then, the analyzer 10 presents recommended values ​​based on the obtained output data (step S202). For example, the analyzer 10 presents recommended setting values ​​based on the setting values ​​output by the model.

[0103] Figure 13 illustrates the flow of the causal discovery process performed by the analysis device 10. Figure 13 is a flowchart showing the flow of the causal discovery process.

[0104] As shown in Figure 13, first, the analyzer 10 extracts neighboring data of the input data from the history DB 122 (step S301). For example, the analyzer 10 extracts K data points in order of their Euclidean distance from the input data.

[0105] Then, the analysis device 10 searches for causality between variables (between features) included in the extracted data (step S302). For example, the analysis device 10 uses LiNGAM to find a linear formula that represents the causal relationship between variables.

[0106] The analysis device 10 presents the obtained causal relationships (step S303). For example, the analysis device 10 may present the causal relationships using a DAG, or it may present them using natural language messages.

[0107] [Effects of the first embodiment] As explained above, the extraction unit 131 extracts a set of neighboring data for a specified data from data, each of which has multiple features. The search unit 134 applies a causal search method to the set of neighboring data to search for causal relationships between features.

[0108] This allows for easy analysis of causal relationships in complex data, according to the embodiment.

[0109] The search unit 134 finds a linear equation that represents the relationship between the features. For example, the search unit 134 represents each feature as the sum of a value obtained by multiplying at least one of the features by a coefficient and noise.

[0110] This allows us to represent local causal relationships between features using linear equations, even when the causal relationships are nonlinear or involve distribution shifts.

[0111] The display control unit 135 outputs the search results from the search unit 134 in the form of natural language text or a directed graph. This allows the user to easily understand the results of the causal relationship analysis.

[0112] [System configuration, etc.] Furthermore, the components of each illustrated device are functionally conceptual and do not necessarily need to be physically configured as shown. In other words, the specific forms of distribution and integration of each device are not limited to those shown, and all or part of them can be functionally or physically distributed or integrated in any unit according to various loads and usage conditions. Moreover, each processing function performed by each device can be implemented, in whole or in any part, by a CPU (Central Processing Unit) and a program that is analyzed and executed by the CPU, or by hardware using wired logic. Note that the program may be executed not only by the CPU but also by other processors such as a GPU.

[0113] Furthermore, among the processes described in the embodiments, 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 by known methods. In addition, the processing procedures, control procedures, specific names, and information including various data and parameters shown in the above document and drawings can be arbitrarily changed unless otherwise specified.

[0114] [program] In one embodiment, the analysis device 10 can be implemented by installing an analysis program that performs the above analysis process as packaged software or online software on a desired computer. For example, by having the above analysis program run on an information processing device, the information processing device can function as the analysis device 10. The information processing device referred to here includes desktop or notebook personal computers. In addition, the category of information processing device also includes tablet terminals, smartphones, mobile phone and PHS (Personal Handyphone System) mobile communication terminals, and slate terminals such as PDA (Personal Digital Assistant).

[0115] Furthermore, the analysis device 10 can also be implemented as a server that provides services related to the analysis process described above to a terminal device used by a user, with the user's terminal device acting as the client. For example, the server may be implemented as a server device that provides analysis services, taking the specification of request points as input and outputting analysis results (e.g., an adjacency matrix). In this case, the server may be implemented as a web server, or it may be implemented as a cloud service that provides the services related to the analysis process described above through outsourcing.

[0116] Figure 14 shows an example configuration of a computer running an analysis program. Computer 1000 has, for example, memory 1010 and a CPU 1020. Computer 1000 also has a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. These components are connected by a bus 1080.

[0117] Memory 1010 includes ROM (Read Only Memory) 1011 and RAM (Random Access Memory) 1012. ROM 1011 stores, for example, a boot program such as BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to the hard disk drive 1090. The disk drive interface 1040 is connected to the disk drive 1100. For example, a removable storage medium such as a magnetic disk or optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to, for example, a mouse 1110 and a keyboard 1120. The video adapter 1060 is connected to, for example, a display 1130.

[0118] The hard disk drive 1090 stores, for example, the OS 1091, application programs 1092, program modules 1093, and program data 1094. That is, the programs defining each process of the analytical device 10 are implemented as program modules 1093 containing code executable by a computer. The program modules 1093 are stored, for example, on the hard disk drive 1090. For example, a program module 1093 for performing processes similar to those in the analytical device 10 is stored on the hard disk drive 1090. Note that the hard disk drive 1090 may be replaced by an SSD (Solid State Drive).

[0119] Furthermore, the configuration data used in the processing of the above-described embodiment is stored as program data 1094 in, for example, memory 1010 or hard disk drive 1090. The CPU 1020 then reads the program module 1093 and program data 1094 stored in memory 1010 or hard disk drive 1090 into RAM 1012 as needed and executes the processing of the above-described embodiment.

[0120] Furthermore, the program module 1093 and program data 1094 are not limited to being stored in the hard disk drive 1090; for example, they may be stored in a removable storage medium and read by the CPU 1020 via a disk drive 1100 or the like. Alternatively, the program module 1093 and program data 1094 may be stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.). The program module 1093 and program data 1094 may then be read by the CPU 1020 from the other computer via a network interface 1070. [Explanation of Symbols]

[0121] 10 Analyzer 11 Communications Department 12 Storage section 13 Control Unit 20a, 20b Plant 21 Terminal device 51 Input data 52 Neighborhood Data 121 Predictive Model Information 122 History DB 131 Extraction part 132 Prediction Section 133 Update Department 134 Search Department 135 Display Control Unit

Claims

1. An extraction unit that extracts a predetermined number of neighbor sets from the accumulated data, in order of proximity between each of the multidimensional vectors whose elements are the multiple feature quantities of each of the accumulated data and the multidimensional vector whose elements are the multiple feature quantities of the required point, A search unit that applies a causal search method to the data of the aforementioned neighborhood set and searches for causality between the plurality of feature quantities, An analytical apparatus characterized by having the following features.

2. The extraction unit extracts the neighborhood sets in order of proximity between each of the multidimensional vectors whose elements are the multiple feature quantities other than time that each of the accumulated data has and the multidimensional vector whose elements are the multiple feature quantities other than time that the requested point has. The analytical apparatus according to claim 1.

3. The analysis apparatus according to claim 1, characterized in that the search unit finds a linear equation representing the relationship between the plurality of feature quantities.

4. The analysis apparatus according to claim 1, characterized in that the search unit represents each of the plurality of features as the sum of a value obtained by multiplying at least one of the plurality of features by a coefficient and noise.

5. The analysis apparatus according to claim 1, further comprising an output unit that outputs the search results from the search unit in the form of natural language text or a directed graph.

6. The search unit searches for causality between the plurality of features by approximating the causal relationship between the plurality of features with a linear model using the data of the neighborhood set. The analytical apparatus according to claim 1.

7. The search unit searches for causality between the plurality of features using LiNGAM. The analytical apparatus according to claim 6.

8. An analytical method performed by an analytical device, An extraction step in which a set of neighbors, which is a predetermined number of data points, is extracted from the accumulated data in order of proximity between each multidimensional vector, which has multiple features as elements, and the multidimensional vector, which has multiple features as elements, that has the required points. A search step in which a causal search method is applied to the data of the aforementioned neighborhood set to search for causality between the multiple feature quantities, An analytical method characterized by including [a certain component].

9. An extraction step of extracting a predetermined number of neighbor sets from the accumulated data, in order of proximity between each of the multidimensional vectors whose elements are the multiple features of each of the accumulated data and the multidimensional vector whose elements are the multiple features of the required point, A search step in which a causal search method is applied to the data of the aforementioned neighborhood set to search for causality between the multiple feature quantities, An analysis program characterized by having a computer execute it.

Images