Information processing device, information processing method, and program

Abstract

To improve the accuracy of estimation using an estimation model for time series data.SOLUTION: An information processing device includes a processing unit. The processing unit adjusts the first distance between first time-series data and a plurality of identification patterns corresponding to a plurality of classes by using the degree of presence that indicates the degree to which a plurality of points in time in the first time-series data is points in time at which one of the plurality of identification patterns is present. The processing unit calculates a first feature quantity representing a feature of the first time-series data by using the adjusted first distance. The processing unit inputs a first feature quantity to an estimation model in which parameters are learned together with similar identification patterns and the degree of presence so as to reduce estimation errors, and thereby obtains identification patterns, classes, and the degree of presence similar to the first time-series data.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] Embodiments of the present invention relate to an information processing apparatus, an information processing method, and a program.

Background Art

[0002] In the classification technology of time series data such as sensor data (time series waveform data), in addition to improving the classification performance (estimation accuracy), it is desirable to clarify the basis of classification. As a classification technology for time series data that clarifies the basis of classification, a supervised shapelet learning method has been proposed in which one or more shapelets, which are short waveform patterns (discrimination patterns) effective for classification, are learned together with a classifier (estimation model), and it has attracted attention in the fields of data mining and machine learning in recent years.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Non-Patent Documents

[0004]

Non-Patent Document 1

Non-Patent Document 2

[0005] The present invention aims to provide an information processing device, an information processing method, and a program that can improve the accuracy of estimation using an estimation model for time series data. [Means for solving the problem]

[0006] The information processing device of this embodiment includes a processing unit. The processing unit adjusts a first distance between the first time series data and multiple identification patterns corresponding to multiple classes using abundance, which represents the degree to which multiple times in the first time series data are times when one of the multiple identification patterns exists. The processing unit calculates a first feature quantity that represents the characteristics of the first time series data using the adjusted first distance. The processing unit inputs the first feature quantity into an estimation model, whose parameters are learned along with similar identification patterns and abundance, in order to reduce estimation errors, thereby obtaining identification patterns, classes, and abundance quantities similar to the first time series data. [Brief explanation of the drawing]

[0007] [Figure 1] Block diagram of the information processing device of the first embodiment. [Figure 2] Flowchart of the learning process in the first embodiment. [Figure 3] A flowchart of the estimation process in the first embodiment. [Figure 4]A diagram showing an example of the display screen of the first embodiment. [Figure 5] A diagram showing an example of a display screen using a comparative example. [Figure 6] Block diagram of the information processing device of the second embodiment. [Figure 7] Flowchart of the learning process in the second embodiment. [Figure 8] Flowchart of the estimation process in the second embodiment. [Figure 9] A diagram showing an example of the display screen of the second embodiment. [Figure 10] Hardware configuration diagram of an information processing device according to the first or second embodiment. [Modes for carrying out the invention]

[0008] A preferred embodiment of the information processing device according to this invention will be described in detail below with reference to the attached drawings.

[0009] The following embodiments can be applied to systems that analyze time-series data, such as sensor data detected and collected by various sensors. The applicable systems are not limited to those described below. A system that classifies time series data into one of several classes. These classes may include, for example, a normal class indicating that the time series data is normal, and an abnormal class indicating that the time series data is abnormal.

[0010] In conventional shapelet learning methods, the distance (similarity) between each shapelet and the most fitting (similar) sub-time series data (sub-waveform) on the time series data is used as a feature. In this case, the region where shapelets exist is basically the entire time series data. The region where shapelets exist can be interpreted as the region of the time series data where shapelets can exist, and is the region that will be matched with shapelets.

[0011] The shapelet learning method can discover shapelets that tolerate positional shifts. However, no technique has yet been proposed to determine which region of time series data a shapelet appears (exists) in. If we were to search for whether a point (time) on the time series data is in the region of existence, we would need to learn the shapelets 2 to the power of Q times, where Q is the length of the time series data (the number of points in the time series data), which is practically impossible from the standpoint of computation time.

[0012] While techniques have been proposed to pre-define the existence domain of shapelets, such techniques cannot discover the existence domain itself.

[0013] In the following embodiment, the existence regions are estimated along with the classification results of the time series data using a classifier (an example of an estimation model) that has been trained to identify multiple existence regions for each shapelet. For example, if the existence region of a shapelet is important for classification, classification performance can be improved. Furthermore, by displaying (visualizing, presenting) the existence regions along with the shapelets that form the basis of the classification, more explainable information can be provided to the user.

[0014] For example, the embodiment has the following functions: • A function that adjusts the distance between a matching shapelet and partial time series data using an abundance level that indicates the degree to which the shapelet exists within that region. • A function to update the presence level so that the presence region of shapelets is obtained in a visually easy-to-understand manner, while reducing the classification loss by the classifier (reducing the estimation error by the estimation model).

[0015] In the following example, we will describe the use of shapelets as the identification pattern for classification, but other patterns besides shapelets may also be used as the identification pattern.

[0016] Here, we will explain the terms and symbols used in the following explanation.

[0017] Hereinafter, let the time-series dataset be T, the number of time-series data included in the time-series dataset T be I, the length of each time-series data be Q points, and the number of shapelets be K. The length of each shapelet is given as L points (where L satisfies L < Q). For simplicity of explanation here, it is assumed that the length of each time-series data is the same Q points, and the length of each shapelet is the same L points. When the length of each time-series data is the same Q points, the time-series dataset T can be considered as an I×Q continuous value matrix. The extension to the case where the lengths are different is easy for each case.

[0018] Also, hereinafter, the case where the time-series data is univariate time-series data will be described as an example. In the case of multivariate time-series data, for example, the method described below can be applied for each variable, and it can be easily extended to identify the shapelets and the existence regions of the shapelets.

[0019] When the length of each shapelet is the same L points, the K shapelets (hereinafter referred to as shapelet S) are a K×L continuous value matrix. The existence regions of the K shapelets are represented by the existence degree M which is a K×J continuous value matrix. J = Q - L + 1, and it represents the length of the region in the length (Q) of the time-series data where the shapelets can be searched. Hereinafter, the existence degree of the k-th shapelet (where k is an integer satisfying 1 ≤ k ≤ K) (hereinafter referred to as shapelet S k is denoted as m k The existence degree m k is a J-dimensional vector. Also, hereinafter, the existence degree (scalar value) of the j-th point (where j is an integer satisfying 1 ≤ j ≤ J) of the k-th shapelet is denoted as m k,j

[0020] The existence degree m k,j For example, the smaller the value, the lower the degree of the existence region of the shapelet. The existence degree may also represent that the larger the value, the lower the degree of the existence region of the shapelet. Hereinafter, the existence degree that represents that the larger the value, the lower the degree of the existence region of the shapelet is referred to as the penalty degree m'.​k,j is as follows. The penalty degree m’ k,j is calculated using the existence degree m k,j by, for example, the following equation (1). The ReLu function in equation (1) is represented by the following equation (2). [Equation] [Equation]

[0021] In equation (1), the penalty degree is calculated to take a value in the range of [1, ∞]. The penalty degree m’ k,j corresponds to the value (scalar value) of the penalty degree when the shapelet S k exists at the j-th point of the time-series data. The penalty degree m’ k for each shapelet S k becomes a J-dimensional vector with J penalty degrees m’ k,j as elements.

[0022] When the value of the penalty degree m’ k,j is small, it is interpreted that the degree to which the j-th point of the time-series data is in the existence region of the shapelet S k is large. When the value of the penalty degree m’ k,j is large, it is interpreted that the degree to which the j-th point of the time-series data is in the existence region of the shapelet S k is small.

[0023] As will be described below, in the embodiment, together with a plurality of shapelets S, the penalty degree m’ k (existence degree m k ) is learned.

[0024] (First Embodiment) FIG. 1 is a block diagram showing an example of the configuration of the information processing apparatus 100 according to the first embodiment. As shown in FIG. 1, the information processing apparatus 100 includes a reception unit 101, an output control unit 102, a distance calculation unit 111, a feature calculation unit 112, a learning unit 113, an estimation unit 114, a storage unit 131, and a display unit 132.

[0025] The reception unit 101 receives input of various types of information used by the information processing device 100. For example, the reception unit 101 receives time-series data used for learning by the learning unit 113, time-series data to be estimated, and specifications of various parameters used in learning or estimation.

[0026] Parameters include information such as the following: • Number of shapelets K • Shapelet length L

[0027] The distance calculation unit 111 calculates the distance DA (first distance) between the time series data TSA (first time series data) input as the target for estimation and the multiple shapelets S. During training, the distance calculation unit 111 calculates the distance DB (second distance) between the time series data TSB (second time series data) input for the training process and the multiple shapelets.

[0028] Here, we have the i-th time series data (where i is an integer satisfying 1 ≤ i ≤ I) and the k-th shapelet S. k The definition of distance is explained below. The distance between time series data and a shapelet is defined as the smallest of J distances calculated by varying the offset j from 1 to J and determining the distance between the extracted partial time series data and the shapelet based on each offset. The offset corresponds to the reference point of the partial time series data included in the time series data that is matched with the shapelet. The reference point is, for example, the starting point (earliest time) of the partial time series data, but it can be any other point, such as the center point or the ending point.

[0029] Thus, shapelets are located at a position that is close (similar) to the partial time series data contained within the time series data. Note that distance can be measured using, for example, the Euclidean distance.

[0030] The distance calculation unit 111 calculates the shapelet s by, for example, the following equation (3): kAnd the i-th time series data t i The partial time series data corresponding to the offset j, and the distance d i,k,j (s k ,t i Calculate L. k represents the length of the k-th shapelet, but if each shapelet is the same length, then L is used for all k. k These have the same value L.

number

[0031] Note that the distance is not limited to the distance calculated by equation (3), and any other type of distance may be used. For example, the distance may be calculated using DTW (Dynamic Time-Warping) distance instead of Euclidean distance.

[0032] The distance calculation unit 111 further calculates the penalty degree m' k The distance (distance DA, DB) is adjusted using the following. In the following, the adjusted distance may be referred to as the adjusted distance. For example, the distance calculation unit 111 applies the penalty degree m' to the distance. k By multiplying by [a certain factor], the adjusted distance is calculated.

[0033] As described above, the penalty degree is calculated to take values ​​in the range [1, ∞]. Therefore, the penalty degree m' is given for distance. k By multiplying by this, the distance calculation unit 111 calculates the adjusted distance such that the adjusted distance increases as the penalty degree increases. As a result, each shapelet is more likely to fit with a point on the time series data with a low penalty degree, and the points on the time series data that fit become the region where the shapelet exists.

[0034] The feature calculation unit 112 calculates feature quantities that represent the characteristics of the time series data using the adjusted distance. For example, the feature calculation unit 112 calculates feature quantity FA (first feature quantity) that represents the characteristics of the time series data TSA using the adjusted distance calculated from the time series data TSA. During training, the feature calculation unit 112 calculates feature quantity FB (second feature quantity) that represents the characteristics of the time series data TSB using the adjusted distance calculated from the time series data TSB.

[0035] Specifically, the feature calculation unit 112 calculates the shapelets s k Using the adjusted distance calculated for each shapelet, k For each, the i-th time series data t i Feature x that represents the characteristics i,k The feature calculation unit 112 calculates the feature quantity x using, for example, the following equation (4). i,k Calculate.

number

[0036] Furthermore, J k This is the k-th shapelet s k This represents the length of the region to search (the range of the offset), but if each shapelet is the same length, then J for all k. k These have the same value J.

[0037] As shown in equation (4), the feature calculation unit 112 calculates features for each shapelet by aggregating the adjusted distances across the entire offset range of the time-series data. Equation (4) shows an example in which the minimum value of multiple adjusted distances corresponding to multiple offsets is used as the feature.

[0038] Note that the distance corresponding to equation (3) can also be considered as dissimilarity. Therefore, similarity may be used instead of distance. For example, similarity can be expressed as a shapelet s k and the i-th time series data t iIt may be defined to take a value of 1 if the partial time series data corresponding to offset j is sufficiently similar, and a value of 0 otherwise.

[0039] Furthermore, instead of taking the minimum value over offset j as in equation (4), the features may be calculated by summing over offset j. In this case, the penalty degree m' for distance is... k Instead of multiplication, use the penalty degree m' for distance. k Division is used. As in this example, features can also be defined using the number of sub-time series data points that are similar to the shapelet on the time series data.

[0040] The learning unit 113 performs the classifier training process using the time series data TSB for training. For example, the learning unit 113 adjusts the classifier parameters and multiple shapelets and penalty m' to reduce the classification loss. k The learning process is executed to update the data.

[0041] For example, the learning unit 113 executes the learning process according to equation (5) below. Equation (5) is given by multiple time series data t i Each loss G i This represents minimizing the classification loss, which is the sum of the classifier parameters w. c,k This represents a set of (weights). c represents C classes (where C is an integer greater than or equal to 2).

number

[0042] (5) Loss G in equation i This is expressed by equation (6) below. Also, y^ in equation (6) i,c This corresponds to the output of the classifier and is expressed by equation (7) below. Note that y^ in equation (6) i,c_i The correct answer is class c. i This corresponds to the classifier output for (c_i). In this way, the learning unit 113 performs a learning process to reduce the error between the class output by the classifier and the correct class.

number

number

[0043] The classifier is not limited to models that output classification results as shown in equation (7), but can be any model that takes features as input to estimate a class, as long as both the shapelets and the penalty (prevalence) parameters can be learned in order to minimize the estimation error.

[0044] Note that the loss G in equation (6) i is the degree of existence m k This includes a second term. The second term corresponds to a term (regularization term) for updating the abundance so that the abundance values ​​of multiple adjacent time points become close together. The second term can be interpreted as a term for updating the abundance so that the region of existence of the shapelet becomes as continuous and wide as possible on the time series data. Note that the loss G without the second term is... i This may be used. If the second clause is not included, the region of existence of the shapelet may not be a region containing multiple consecutive points.

[0045] The learning unit 113 reduces the classification loss by, for example, following equation (5), by changing the shapelet S and the classifier parameter w c,k The shapelet classifier update unit is used. The method for updating the classifier can be any conventional method, but for example, a method can be applied in which the parameter values ​​are slightly updated in the direction that reduces the classification loss using gradient descent (see, for example, Non-Patent Document 2).

[0046] Furthermore, the learning unit 113 updates the penalty rate for each shapelet at points fitted on the time-series data to reduce the classification loss (penalty rate update unit). The method for updating the penalty rate can be the same as for the parameters of the shapelet S and the classifier, for example, a method using gradient descent can be applied.

[0047] The estimation unit 114 uses the classifier trained by the learning unit 113 to perform estimation processing on the time series data TSA input as the target of estimation. For example, the estimation unit 114 inputs the feature quantity FA calculated for the time series data TSA to the trained classifier, and obtains the class estimated by the classifier, the shapelet corresponding to the class, and the penalty degree m'. k And, we seek.

[0048] The classifier is calculated by, for example, equation (7) y^ i,c The class c with the highest value may be output as the classification result, or the probability of each class c may be output as the classification result. The probability of each class c can be calculated, for example, by the formula listed as the argument to the log function in equation (6).

[0049] The output control unit 102 controls the output of various types of information used by the information processing device 100. For example, the output control unit 102 outputs (displays) information indicating the estimation results from the estimation unit 114 to the display unit 132. The estimation results include, for example, the shapelet output for the feature quantity FA, ​​the class which is the classification result, and the penalty degree m'. k and, including. Penalty degree m' k This corresponds to information indicating whether or not it is a domain of existence.

[0050] At least a portion of each of the above-mentioned parts (reception unit 101, output control unit 102, distance calculation unit 111, feature calculation unit 112, learning unit 113, and estimation unit 114) may be implemented by one or more processing units. Each of the above-mentioned parts may be implemented by, for example, one or more processors. For example, each of the above-mentioned parts may be implemented by having a processor such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit) execute a program, i.e., by software. Each of the above-mentioned parts may be implemented by a processor such as a dedicated IC (Integrated Circuit), i.e., by hardware. Each of the above-mentioned parts may be implemented by using a combination of software and hardware. When multiple processors are used, each processor may implement one of the above-mentioned parts, or two or more of the above-mentioned parts.

[0051] The memory unit 131 stores various types of information used in the information processing device. For example, the memory unit 131 stores the time series data TSB used for learning, the time series data TSA to be estimated, and various parameters (such as the number of shapelets K and the length of the shapelets L). The memory unit 131 also stores the classifier parameters (such as weights), shapelets, and penalty levels obtained through the learning process.

[0052] The storage unit 131 can be composed of any commonly used storage medium, such as flash memory, memory card, RAM (Random Access Memory), HDD (Hard Disk Drive), and optical disc.

[0053] The display unit 132 is an example of a device that displays various types of information used by the information processing device 100. The display unit 132 is implemented by a display device such as a liquid crystal display.

[0054] The information processing device 100 may be composed of one physical device or multiple physical devices. For example, the information processing device 100 may be built on a cloud environment. Also, each part of the information processing device 100 may be distributed and provided on multiple devices. For example, the information processing device 100 (information processing system) may be configured to include a device (e.g., a learning device) equipped with functions necessary for learning (e.g., a learning unit 113) and a device (e.g., an estimation device) equipped with functions necessary for estimation (e.g., an estimation unit 114).

[0055] Next, the learning process flow of the information processing device 100 of the first embodiment will be described. Figure 2 is a flowchart showing an example of the learning process in the first embodiment. The purpose of the learning process is to identify one or more shapelets that are effective waveform patterns (identification patterns) for classification, the region of existence (penalty level) of each shapelet in the time-series data, and the parameters of the classifier.

[0056] The reception unit 101 accepts input of a time series dataset containing training time series data TSB (step S101). The time series data TSB is data to which labels indicating the correct class are assigned. The reception unit 101 may also accept input of parameters such as the number of shapelets K and the length of the shapelets L. It may also be configured not to accept input of parameters and to use default values. The default value for the number of shapelets K is, for example, 20. The default value for the length of the shapelets L is calculated, for example, by L = Q × (value indicating a proportion). The value indicating a proportion may be set to a fixed value such as 0.1.

[0057] The learning unit 113 initializes the shapelet S, the penalty degree, and the classifier parameters (step S102). The simplest initialization method is to set random values.

[0058] Subsequently, the parameters for the shapelet S, penalty degree, and classifier are learned by repeating the following steps (steps S103 to S109).

[0059] First, the distance calculation unit 111 calculates the distance between each of the I time series data points and the K shapelets (step S103). The distance calculation unit 111 then calculates the adjusted distance using the penalty degree of each point in the time series data (step S104).

[0060] The feature calculation unit 112 calculates the feature quantity from the adjusted distance using, for example, equation (4) (step S105).

[0061] The learning unit 113 calculates the classification loss from the calculated features and the classifier parameters. The learning unit 113 then updates the shapelets and the classifier parameters to reduce the classification loss (step S106). The learning unit 113 also updates the penalty rate to reduce the classification loss (step S107). Furthermore, the learning unit 113 updates the penalty rate so that the penalty rates of adjacent points approach the same value (step S108).

[0062] The learning unit 113 determines whether or not to terminate the loop (step S109). For example, the learning unit 113 determines to terminate the loop when the number of repetitions reaches the upper limit. If it is determined not to terminate the loop (step S109: No), the process returns to step S103 and is repeated.

[0063] If it is determined that the iteration should be terminated (step S109: Yes), the learning process ends. The learning unit 113 stores the shapelets S, penalty degree, and classifier parameters obtained through the learning process in the storage unit 131.

[0064] Next, the flow of the estimation process by the information processing device 100 of the first embodiment will be described. Figure 3 is a flowchart showing an example of the estimation process in the first embodiment. The purpose of the estimation process is to classify the time series data TSA to be estimated using the classifier (estimation model), shapelets, and penalty degree learned through the learning process, and to visualize the classification result, the shapelets that contributed to the classification, and the region in which the shapelets exist, together with the time series data.

[0065] The reception unit 101 accepts input of the time series data TSA to be estimated, as well as information on the trained classifier (estimation model), shapelets, and penalty rates (step S201). In the estimation process, the parameters of the shapelets S, penalty rates, and classifiers are not initialized as during training; the input information is used as is.

[0066] Steps S202 to S204 for calculating features for time series data are the same as steps S103 and S105 of the learning process.

[0067] The estimation unit 114 performs classification using the calculated features and classifier parameters (step S205).

[0068] The output control unit 102 outputs the estimation results from the estimation unit 114 (step S206) and terminates the estimation process. For example, the output control unit 102 displays a display screen on the display unit 132 that visualizes the shapelets that contributed to the classification and the regions in which those shapelets exist, along with time-series data.

[0069] Figure 4 shows an example of the display screen of the first embodiment. Figure 4 shows an example of visualization when the number of shapelets 402 that contributed to the classification is 1, and the region 403 in which that shapelet 402 exists is identified as a single region near the center of the time series data 401. The bar graph at the bottom of Figure 4 shows an example of the distribution of features (distance from shapelets) when classification of a large number of time series data is performed.

[0070] The existence region 403 is identified, for example, as a point (time) where the penalty degree is below a threshold. By learning using a loss that includes a term to update the existence (penalty degree) values ​​of multiple adjacent time points to be close together, as in equation (6), the existence region can be made as wide and continuous as possible on the time series data.

[0071] Figure 5 shows an example of the display screen in a comparative example where the existence region of the shapelets is not learned. Figure 5 shows an example of visualizing shapelets 502 on the time series data 501 when the number of shapelets 502 that contributed to the classification was 1. For the sake of explanation, the existence region 503, which corresponds to the entire area of ​​the time series data 501, is displayed, but the existence region 503 does not necessarily have to be visualized.

[0072] Note that in the time-series data of Figures 4 and 5, the slopes of the peaked waveforms near the center are different, but the slopes of the peaked waveforms near both ends are the same. In other words, Figures 4 and 5 are examples of how partial time-series data near the center can be classified into one of two classes (e.g., a normal class and an abnormal class).

[0073] In the comparative example where the existence region is not learned (Figure 5), the shapelet can fit any region on the time series data. Therefore, for example, the shapelet may be incorrectly fitted to time series data with superimposed noise. Figure 5 shows an example where the shapelet was incorrectly fitted to the waveform near the left edge instead of the waveform near the center. As a result, the shapelet is not correctly learned, and the estimation process cannot sufficiently separate the features between the two classes for the time series data being estimated. The bar graph at the bottom of Figure 5 shows how the feature distribution is not separated.

[0074] In contrast, in this embodiment (Figure 4), which learns the region of existence, the region of existence of shapelets is identified while learning the shapelets. In the example in Figure 4, the region of existence excluding the areas near both ends is learned. As a result, even in time-series data with superimposed noise, the shapelets fit the waveform in the center. Consequently, the shapelets are also learned correctly, and the features between the two classes can be separated more appropriately in the time-series data to be estimated. Furthermore, the user can be presented with the information that shapelets effective for classification are located near the center of the time-series data and are not present at the ends.

[0075] Thus, in the first embodiment, in addition to training the shapelets and classifier, one or more regions on the time series data in which each shapelet exists are simultaneously and efficiently trained. This improves the accuracy of estimation using the estimation model on the time series data. Furthermore, by displaying (visualizing, presenting) not only the shapelets that form the basis of the classification but also the regions in which they exist, it becomes possible to provide users with more explainable information.

[0076] (Second embodiment) The distance between each shapelet and the partial time series data may include distances of multiple types (components), such as distances related to the waveform shape, distances related to the waveform height, and distances related to the waveform magnitude. Conventional shapelet learning methods do not take into account the use of such multiple types of distances, nor the degree to which these multiple types of distances contribute to classification (hereinafter referred to as "contribution").

[0077] The information processing device of the second embodiment classifies time-series data using multiple types of distance and visualizes the contribution of each of the multiple types of distance to the classification.

[0078] Figure 6 is a block diagram showing an example of the configuration of the information processing device 100-2 according to the second embodiment. As shown in Figure 6, the information processing device 100-2 includes a reception unit 101, an output control unit 102-2, a distance calculation unit 111-2, a feature calculation unit 112-2, a learning unit 113-2, an estimation unit 114, a storage unit 131, and a display unit 132.

[0079] In the second embodiment, the functions of the output control unit 102-2, distance calculation unit 111-2, feature calculation unit 112-2, and learning unit 113-2 differ from those of the first embodiment. The other configurations and functions are the same as those in Figure 1, which is a block diagram of the information processing device 100 of the first embodiment, so they are denoted by the same reference numerals and their description is omitted here.

[0080] The distance calculation unit 111-2 calculates the distance (distance DA, distance DB) between the time series data and multiple shapelets. First, the distance calculation unit 111-2 calculates several types of distances. Hereinafter, each of the several types of distances may be referred to as a distance component. The distance calculation unit 111-2 calculates distance DA and distance DB by integrating the multiple distance components using contributions. As a method of integration using contributions, for example, a method can be applied in which the average value obtained by weighting each distance component by its contribution is calculated as distance DA or distance DB.

[0081] Examples of different distance types are shown below. The difference between the statistics of the partial time series data and the statistics of the shapelet. The statistics are, for example, the mean, standard deviation, or variance of the values ​​at each point. The distance between the shape of the partial time series data and the shape of the shapelet. The distance between the shape of the standardized partial time series data and the shape of the shapelet. The distance between the frequency components of the partial time series data and the frequency components of the shapelet. The distance between the p-th derivative (where p is an integer greater than or equal to 1) of a partial time series data and the p-th derivative of a shapelet.

[0082] The difference in mean values ​​can be interpreted as the shift in the height direction of the sensor values ​​(the height direction of the waveform). The difference in standard deviations or variances can be interpreted as the difference in amplitude. The distance between shapes can be either Euclidean distance or DTW (Dynamic Time-Warping) distance. Standardization means, for example, eliminating the shift in the height direction of the sensor values ​​(i.e., setting the mean value of the partial time series data to 0) and equalizing the amplitude (i.e., setting the standard deviation of the partial time series data to 1).

[0083] In this embodiment, two or more types of distances (distance components) from the multiple types of distances described above are used. Hereinafter, the number of distance components will be N (where N is an integer greater than or equal to 2). Each of the K shapelets S k Since there are N contributions as parameters, the contribution matrix V is a continuous matrix of K × N values. Below, the nth contribution (where n is an integer satisfying 1 ≤ n ≤ N) is given by v n It is sometimes expressed as follows: Contribution v n This is a K-dimensional vector.

[0084] For example, the distance calculation unit 111-2 calculates the distance d by weighting multiple types of distances (distance components) by their contribution using the following equation (8). i,k,j (s k ,t i Calculate (8). Note that equation (8) is an example for the case where N=3.

number

[0085] (8) t in equation normalized This is standardized partial time series data, which can be calculated, for example, by equation (9) below.

number

[0086] (8) s in equation meam and s stdThese are the mean and variance of the shapelet, and are calculated, for example, by equations (10) and (11) below.

number

number

[0087] (8) t in equation meam and s std These are the mean and variance of the partial time series data, and are calculated, for example, by equations (12) and (13) below.

number

number

[0088] (8) v in equation n This is the normalized contribution v n It can be calculated, for example, using equation (14) below.

number

[0089] The feature calculation unit 112-2 calculates the feature quantity x using, for example, the following equation (15): i,k Equation (15) calculates the feature quantity x of the first embodiment, using the contribution V and the distance calculated by, for example, equation (8). i,k This is different from equation (4) which is used to calculate [the result].

number

[0090] The learning unit 113-2 reduces the classification loss by adjusting the classifier parameters and multiple shapelets and penalty m'. k The learning process is executed to update the contribution V and .

[0091] For example, the learning unit 113-2 performs the learning process according to equation (16) below. The loss G in equation (16) i This is the same as equation (6) above. In this way, the learning unit 113-2 performs a learning process that further includes updating the contribution score in order to reduce the error between the class output by the classifier and the correct class. The method for updating the contribution score V can be the same as for the shapelet S, classifier parameters and penalty score, for example, a method using gradient descent.

number

[0092] The output control unit 102-2 outputs the shapelet, class, and penalty degree m' obtained by the estimation process. k In addition, it differs from the output control unit 102 of the first embodiment in that it further outputs the contribution level.

[0093] Next, the learning process by the information processing device 100-2 of the second embodiment will be described with reference to Figure 7. Figure 7 is a flowchart showing an example of the learning process in the second embodiment. In this embodiment, in addition to the shapelet S, penalty degree, and classifier parameters, the contribution V of the distance component is learned.

[0094] The reception unit 101 accepts input of a time series dataset including training time series data TSB, similar to step S101 (Figure 2) of the first embodiment (step S301).

[0095] The learning unit 113-2 initializes the shapelet S, the penalty degree, and the classifier parameters, as well as the contribution (step S302). The simplest initialization method is to set a random value.

[0096] Subsequently, the following steps (steps S303 to S311) are repeated to learn the shapelet S, penalty degree, classifier parameters, and contribution.

[0097] The distance calculation unit 111-2 calculates multiple types of distances between I time series data and K shapelets (step S303). The distance calculation unit 111-2 integrates the multiple types of distances using the contribution (step S304). The distance calculation unit 111-2 calculates a distance weighted by the contribution of multiple types of distances (distance components) using, for example, equation (8).

[0098] Steps S305 to S308 are the same as steps S104 to S107 in the information processing device 100 of the first embodiment, so their explanation will be omitted.

[0099] In this embodiment, the learning unit 113-2 updates the contribution of each distance component to reduce the classification loss (step S309).

[0100] Steps S310 to S311 are the same as steps S108 to S109 in the information processing device 100 of the first embodiment, so their explanation will be omitted.

[0101] The learning unit 113-2 stores the shapelets S, penalty degree, classifier parameters, and contributions obtained through the learning process in the storage unit 131.

[0102] Next, the estimation process by the information processing device 100-2 of the second embodiment will be explained using Figure 8. Figure 8 is a flowchart showing an example of the estimation process in the second embodiment. In this embodiment, the classification result, the shapelets that contributed to the classification, and the region in which the shapelets exist, as well as the degree of contribution, are visualized.

[0103] The reception unit 101 accepts input of the time series data TSA to be estimated, as well as the trained classifier (estimation model), shapelets, and penalty information, similar to step S201 (Figure 3) of the first embodiment (step S401).

[0104] Steps S402 to S405 for calculating features for time series data are the same as steps S303 and S306 of the learning process in Figure 7.

[0105] The estimation unit 114 performs classification using the calculated features and classifier parameters (step S406).

[0106] The output control unit 102-2 outputs the estimation results from the estimation unit 114 (step S407) and terminates the estimation process. For example, the output control unit 102-2 displays on the display unit 132 a display screen that visualizes the shapelets that contributed to the classification, the regions in which the shapelets that contributed to the classification exist, and the degree of contribution in relation to time-series data.

[0107] Figure 9 shows an example of the display screen of the second embodiment. Figure 9 is an example of a display screen in which a display area 901 for displaying the contribution is added to the display screen of the first embodiment (Figure 4). As shown in Figure 9, the display of the contribution makes it possible to understand the distance component that contributes to the classification for each shapelet. In the example of Figure 9, it can be seen that the difference due to the shape of the waveform is more effective for classification than the difference in the mean and variance.

[0108] Thus, the information processing device of the second embodiment can identify and visualize the contribution of each distance component, such as shape deviation, height deviation (average) of sensor values, and amplitude magnitude (variance) deviation. This allows the user to utilize the contribution as a basis for classification decisions.

[0109] As described above, the first and second embodiments make it possible to improve the accuracy of estimation using estimation models for time series data.

[0110] Next, the hardware configuration of the information processing device according to the first or second embodiment will be described using Figure 10. Figure 10 is an explanatory diagram showing an example of the hardware configuration of the information processing device according to the first or second embodiment.

[0111] The information processing device in the first or second embodiment includes a control device such as a CPU 51, a storage device such as a ROM (Read Only Memory) 52 or RAM 53, a communication I / F 54 for communication via a network, and a bus 61 for connecting the various parts.

[0112] The program to be executed in the information processing device of the first or second embodiment is provided pre-installed in a ROM 52 or the like.

[0113] The program executed by the information processing device of the first or second embodiment may be configured to be provided as a computer program product by recording it in an installable or executable file format onto a computer-readable recording medium such as a CD-ROM (Compact Disk Read Only Memory), a flexible disk (FD), a CD-R (Compact Disk Recordable), or a DVD (Digital Versatile Disk).

[0114] Furthermore, the information processing device of the first or second embodiment may be configured to store the program on a computer connected to a network such as the Internet and provide it by allowing download via the network. Alternatively, the information processing device of the first or second embodiment may be configured to provide or distribute the program via a network such as the Internet.

[0115] A program executed in the information processing device of the first or second embodiment can cause a computer to function as a part of the information processing device described above. This computer can read a program from a computer-readable storage medium onto its main memory and execute it using the CPU 51.

[0116] An example of the configuration of the embodiment is described below. (Configuration Example 1) The first distance between the input first time series data and the multiple identification patterns corresponding to each of the multiple classes is adjusted using an abundance score that represents the degree to which the multiple time points in the first time series data are the time points in which one of the multiple identification patterns exists. Using the adjusted first distance, a first feature quantity representing the characteristics of the first time series data is calculated. A model that takes a second feature representing the characteristics of a second time series data for training as input and estimates the class corresponding to a classification pattern similar to the second time series data among a plurality of classification patterns, wherein the estimation model learns parameters for similar classification patterns and the frequency of existence of a similar classification pattern at a given time, and by inputting the first feature into the estimation model, the model obtains the classification pattern similar to the first time series data, the class, and the frequency of existence. Processing unit, An information processing device equipped with the following features. (Configuration example 2) The aforementioned processing unit, Using the second time series data, a learning process is performed to update the parameters of the estimation model, the multiple identification patterns, and the presence of each pattern in order to reduce the error between the class estimated by the estimation model and the correct class. The information processing device described in Configuration Example 1. (Configuration Example 3) The aforementioned processing unit, The presence values ​​are updated so that the presence values ​​of multiple adjacent time points become close to each other. The information processing device described in Configuration Example 2. (Configuration example 4) The aforementioned processing unit, The first distance is adjusted by multiplying it by the aforementioned abundance. An information processing device as described in any one of Configuration Examples 1 to 3. (Configuration example 5) The first distance is the sum of the distances between the first time series data and one or more of the identification patterns from the plurality of identification patterns. An information processing device as described in any one of Configuration Examples 1 to 4. (Configuration example 6) The aforementioned processing unit, The class estimated for the first feature, the identification pattern corresponding to the estimated class, and the abundance are displayed on the display device. An information processing device as described in any one of Configuration Examples 1 to 5. (Configuration example 7) The first distance is calculated based on multiple types of distances between the first time series data and multiple identification patterns, and the contribution of each of the multiple types of distances. The estimation model learns the parameters of similar identification patterns, the frequency of occurrence of similar identification patterns at a given time, and the contribution of those patterns, in order to minimize estimation errors. An information processing device as described in any one of Configuration Examples 1 to 6. (Configuration example 8) Multiple types of distances are, The difference between the statistics of the partial time series data corresponding to the length of the identification pattern included in the first time series data and the statistics of the identification pattern, The distance between the shape of the partial time series data and the shape of the identification pattern, The distance between the shape of the standardized partial time series data and the shape of the identification pattern, The distance between the frequency components of the partial time series data and the frequency components of the identification pattern, The distance between the p-th derivative (where p is an integer greater than or equal to 1) of the aforementioned partial time series data and the p-th derivative of the aforementioned identification pattern, Including two or more of the following: The information processing device described in Configuration Example 7. (Configuration example 9) The aforementioned processing unit, The class estimated for the first feature, the identification pattern corresponding to the estimated class, the abundance, and the contribution are displayed on the display device. The information processing device described in Configuration Example 7. (Configuration example 10) The aforementioned processing unit, A distance calculation unit that adjusts the first distance using the abundance, A feature calculation unit that calculates the aforementioned first feature quantity, An estimation unit that inputs the first feature quantities into the estimation model to obtain the identification pattern, the class, and the abundance that are similar to the first time series data. Equipped with, An information processing device as described in any one of Configuration Examples 1 to 9. (Configuration Example 11) An information processing method performed by an information processing device, The first step of adjusting the first distance between the input first time series data and multiple identification patterns corresponding to each of the multiple classes using an abundance score that represents the degree to which multiple times in the first time series data are times when one of the multiple identification patterns exists. A step of calculating a first feature quantity that represents the characteristics of the first time series data using the adjusted first distance, A model that takes a second feature representing the characteristics of a second time series data for training as input and estimates the class corresponding to a classification pattern similar to the second time series data among a plurality of classification patterns, wherein the parameters of the similar classification pattern and the presence of the similar classification pattern at a given time are learned in order to minimize the estimation error, and the first feature is input to the estimation model to obtain the classification pattern, the class, and the presence of the first time series data. Information processing methods including (Configuration Example 12) On the computer, The first step of adjusting the first distance between the input first time series data and multiple identification patterns corresponding to each of the multiple classes using an abundance score that represents the degree to which multiple times in the first time series data are times when one of the multiple identification patterns exists. A step of calculating a first feature quantity that represents the characteristics of the first time series data using the adjusted first distance, A model that takes a second feature representing the characteristics of a second time series data for training as input and estimates the class corresponding to a classification pattern similar to the second time series data among a plurality of classification patterns, wherein the parameters of the similar classification pattern and the presence of the similar classification pattern at a given time are learned in order to minimize the estimation error, and the first feature is input to the estimation model to obtain the classification pattern, the class, and the presence of the first time series data. A program to execute.

[0117] While several embodiments of the present invention have been described, these embodiments are presented as examples only and are not intended to limit the scope of the invention. These novel embodiments can be carried out in a variety of other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their variations are included in the scope and spirit of the invention, as well as in the claims of the invention and its equivalents. [Explanation of symbols]

[0118] 100, 100-2 Information Processing Device 101 Reception Department 102, 102-2 Output control unit 111, 111-2 Distance calculation unit 112, 112-2 Feature Calculation Unit 113, 113-2 Learning Department 114 Estimation Department 131 Storage section 132 Display section

Claims

1. The first distance between the input first time series data and the multiple identification patterns corresponding to each of the multiple classes is adjusted using an abundance score that represents the degree to which the multiple time points in the first time series data are the time points in which one of the multiple identification patterns exists. Using the adjusted first distance, a first feature quantity representing the characteristics of the first time series data is calculated. A model that takes a second feature representing the characteristics of a second time series data for training as input and estimates the class corresponding to a classification pattern similar to the second time series data among a plurality of classification patterns, wherein the estimation model learns parameters for similar classification patterns and the frequency of existence of a similar classification pattern at a given time, and by inputting the first feature into the estimation model, the model obtains the classification pattern, the class, and the frequency of existence that are similar to the first time series data. Processing unit, An information processing device equipped with the following features.

2. The aforementioned processing unit, Using the second time series data, a learning process is performed to update the parameters of the estimation model, the multiple identification patterns, and the presence of each pattern in order to reduce the error between the class estimated by the estimation model and the correct class. The information processing apparatus according to claim 1.

3. The aforementioned processing unit, The presence values ​​are updated so that the presence values ​​of multiple adjacent time points become close to each other. The information processing apparatus according to claim 2.

4. The aforementioned processing unit, The first distance is adjusted by multiplying it by the aforementioned abundance. The information processing apparatus according to claim 1.

5. The first distance is the sum of the distances between the first time series data and one or more of the identification patterns from the plurality of identification patterns. The information processing apparatus according to claim 1.

6. The aforementioned processing unit, The class estimated for the first feature, the identification pattern corresponding to the estimated class, and the abundance are displayed on the display device. The information processing apparatus according to claim 1.

7. The first distance is calculated based on multiple types of distances between the first time series data and multiple identification patterns, and the contribution of each of the multiple types of distances. The estimation model learns the parameters of similar identification patterns, the frequency of occurrence of similar identification patterns at a given time, and the contribution of those patterns, in order to minimize estimation errors. The information processing apparatus according to claim 1.

8. Multiple types of distances are, The difference between the statistics of the partial time series data corresponding to the length of the identification pattern included in the first time series data and the statistics of the identification pattern, The distance between the shape of the partial time series data and the shape of the identification pattern, The distance between the shape of the standardized partial time series data and the shape of the identification pattern, The distance between the frequency components of the partial time series data and the frequency components of the identification pattern, The distance between the p-th derivative (where p is an integer greater than or equal to 1) of the partial time series data and the p-th derivative of the identification pattern, Including two or more of the following: The information processing apparatus according to claim 7.

9. The aforementioned processing unit, The class estimated for the first feature, the identification pattern corresponding to the estimated class, the abundance, and the contribution are displayed on the display device. The information processing apparatus according to claim 7.

10. The aforementioned processing unit, A distance calculation unit that adjusts the first distance using the abundance, A feature calculation unit that calculates the first feature quantity, An estimation unit that inputs the first feature quantities into the estimation model to obtain the identification pattern, the class, and the abundance that are similar to the first time series data. Equipped with, The information processing apparatus according to claim 1.

11. An information processing method performed by an information processing device, The first step of adjusting the first distance between the input first time series data and a plurality of identification patterns corresponding to each of a plurality of classes using an abundance score that represents the degree to which a plurality of times in the first time series data is a time when one of the plurality of identification patterns exists. A step of calculating a first feature quantity that represents the characteristics of the first time series data using the adjusted first distance, A model that takes a second feature representing the characteristics of a second time series data for training as input and estimates the class corresponding to an identification pattern similar to the second time series data among a plurality of identification patterns, wherein the parameters of the similar identification pattern and the existence of the similar identification pattern at a given time are learned in order to minimize estimation errors, and the first feature is input to the estimation model to obtain the identification pattern, the class, and the existence of the first time series data. Information processing methods including

12. On the computer, The first step of adjusting the first distance between the input first time series data and a plurality of identification patterns corresponding to each of a plurality of classes using an abundance score that represents the degree to which a plurality of times in the first time series data is a time when one of the plurality of identification patterns exists. A step of calculating a first feature quantity that represents the characteristics of the first time series data using the adjusted first distance, A model that takes a second feature representing the characteristics of a second time series data for training as input and estimates the class corresponding to an identification pattern similar to the second time series data among a plurality of identification patterns, wherein the parameters of the similar identification pattern and the existence of the similar identification pattern at a given time are learned in order to minimize estimation errors, and the first feature is input to the estimation model to obtain the identification pattern, the class, and the existence of the first time series data. A program to execute.

Images