Information processing device, information processing method, and program.

Abstract

To provide a technique for predicting time-series data indicating a semi-periodic pattern which can estimate abnormality detection and an individual period.SOLUTION: A control part 11 of an information processor 10 acquires time-series data indicating a semi-periodic pattern in a predetermined period, generates a feature vector with a feature amount for each channel as an element for each of a plurality of time points, on the basis of the acquired time-series data using a convolution neutral network to which a kernel having a plurality of channels is applied, weights the feature vectors at the plurality of time points on the basis of similarity between the feature vectors at the plurality of time points and a feature vector at a first time point among the plurality of time points using an attention mechanism to generate an output vector of the attention mechanism, and inputs the output vector to a whole coupling layer of the convolution neutral network to predict time-series data after the predetermined period.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to an information processing device, an information processing method, and a program. [Background technology]

[0002] Industry 4.0, a new industrial stage that focuses on the integration of vertical and horizontal manufacturing processes and the connection of things to the Internet, is attracting attention. The smart factories advocated by Industry 4.0 facilitate manufacturing processes and enable efficient production through communication between all machines and systems within the factory. In production lines where each production machine is equipped with a smart meter to measure power consumption, if abnormal signals can be automatically detected from fluctuations in the time-series data displayed by the smart meter, product defect rates can be reduced and productivity can be improved.

[0003] For example, Patent Document 1 discloses a technology for detecting abnormalities that may be a sign of failure in equipment installed in a factory or the like, using machine learning techniques such as a support vector machine, a kNN (k-nearest neighbor) algorithm, a k-means algorithm, a neural network, or a local outlier factor method. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] International Publication No. 2018 / 003879 Summary of the Invention [Problem to be solved by the invention]

[0005] The technology disclosed in Patent Document 1 has room for improvement in the following respects when detecting anomalies in time-series data that exhibit quasi-periodic patterns, such as power consumption data of production machinery.

[0006] That is, as shown in Figure 1 below, time-series data from a smart meter on a production line shows a quasi-periodic pattern, but does not include system information such as lot changeovers or shutdown states. For such time-series data, there is a strong demand for estimation of individual periods, which are important product performance indicators, in parallel with anomaly detection. This is because estimating individual periods makes it possible to estimate the degree of anomaly on the production line and detect lot changeovers. There are two main approaches to anomaly detection for time-series data showing quasi-periodic patterns, such as power consumption data from production machinery.

[0007] The first approach is a segment-based method that divides time series data into segments equivalent to one cycle and then classifies them using the features of the segments. In the first stage of segmentation, characteristic candidate points (e.g., spikes or other points that appear once per cycle) that appear within a cycle are extracted, and the data is divided into segments using these candidate points. In the second stage of anomaly detection, classification methods such as naive Bayes or support vector machines are used when teacher labels indicating anomalies or normality are provided. In the second stage, clustering methods such as density-based spatial clustering of applications with noise (DBSCAN) or k-means are used when teacher labels are not provided. Furthermore, unsupervised anomaly detection methods such as kNN or support vector data description (SVDD), or methods utilizing deep learning, have also been proposed, taking advantage of the fact that the majority of data is normal. However, the former classification method is difficult to apply to power consumption data from production machinery because teacher labels are not provided and assigning teacher labels is difficult from the perspective of reducing labor. Even with the latter clustering method, high accuracy is required for the first-stage segmentation and the second-stage clustering in anomaly detection, but it is difficult to return a highly accurate response immediately.

[0008] The second approach is a forecasting-based method that utilizes time-series data forecasting techniques to evaluate the degree of anomaly based on the forecast error, enabling immediate response. Representative forecasting-based methods include research using the ARIMA (Autoregressive Integrated Moving Average) model or LSTM (Long Short Term Memory), which distinguish anomalies inside and outside the prediction confidence interval. However, these methods cannot estimate individual cycles, nor can they detect anomalies using quasi-periodic patterns.

[0009] In view of the above circumstances, an object of the present invention is to provide a technique for predicting time series data showing quasi-periodic patterns, which enables anomaly detection and individual periodicity estimation. [Means for solving the problem]

[0010] An information processing device according to an embodiment of the present invention includes: An information processing device including a control unit, The control unit a time series data acquisition process for acquiring time series data showing a quasi-periodic pattern over a predetermined period; a feature vector generation process for generating, for each of a plurality of time points, a feature vector having a feature amount for each of the channels as an element based on the acquired time series data, using a convolutional neural network to which a kernel having a plurality of channels is applied; an output vector generation process that uses an attention mechanism to weight the feature vectors at the plurality of time points based on a similarity between the feature vectors at the plurality of time points and the feature vector at the earliest time point among the plurality of time points, thereby generating an output vector from the attention mechanism; a time series data prediction process for predicting the time series data after the predetermined period by inputting the weighted feature vectors at the plurality of time points into a fully connected layer of the convolutional neural network; Execute.

[0011] In addition, in an information processing device according to an embodiment of the present invention, The control unit may estimate an individual period at each time of the acquired time series data based on a time point among the plurality of time points, excluding the earliest time point, at which a value of an element of a weight vector corresponding to the similarity is maximum.

[0012] In addition, in an information processing device according to an embodiment of the present invention, The control unit may determine the presence or absence of an abnormality in the acquired time series data based on a comparison between the acquired time series data and the predicted time series data.

[0013] In addition, in an information processing device according to an embodiment of the present invention, The control unit may determine the presence or absence of an abnormality in the acquired time-series data based on a probability distribution configured using elements of a weight vector corresponding to the degree of similarity.

[0014] In addition, in an information processing device according to an embodiment of the present invention, The control unit may determine the presence or absence of an abnormality in the acquired time-series data based on an empirical distribution constructed using the estimated frequencies of the individual periods.

[0015] An information processing method according to an embodiment of the present invention includes: An information processing method executed by an information processing device, a time-series data acquisition step of acquiring time-series data showing a quasi-periodic pattern for a predetermined period; a feature vector generation step of generating, for each of a plurality of time points, a feature vector having a feature amount for each of the channels as an element based on the acquired time series data by using a convolutional neural network to which a kernel having a plurality of channels is applied; an output vector generation step of generating an output vector of the attention mechanism by weighting the feature vectors at the plurality of time points based on a similarity between the feature vectors at the plurality of time points and the feature vector at the earliest time point among the plurality of time points; a time series data prediction step of predicting the time series data after the predetermined period by inputting the weighted feature vectors at the plurality of time points into a fully connected layer of the convolutional neural network; Includes:

[0016] A program according to an embodiment of the present invention includes: On the computer, a time-series data acquisition step of acquiring time-series data showing a quasi-periodic pattern for a predetermined period; a feature vector generation step of generating, for each of a plurality of time points, a feature vector having a feature amount for each of the channels as an element based on the acquired time series data by using a convolutional neural network to which a kernel having a plurality of channels is applied; an output vector generation step of generating an output vector of the attention mechanism by weighting the feature vectors at the plurality of time points based on a similarity between the feature vectors at the plurality of time points and the feature vector at the earliest time point among the plurality of time points; a time series data prediction step of predicting the time series data after the predetermined period by inputting the weighted feature vectors at the plurality of time points into a fully connected layer of the convolutional neural network; Execute the following. [Effects of the Invention]

[0017] According to the present invention, it is possible to provide a technique for predicting time series data exhibiting quasi-periodic patterns, which enables anomaly detection and estimation of individual periods. [Brief explanation of the drawings]

[0018] [Figure 1] FIG. 1 is a diagram illustrating an outline of time-series data of a smart meter in a certain production line. [Figure 2] 1 is a block diagram showing a schematic configuration of an information processing device 10 according to an embodiment of the present invention. [Figure 3] 3 is a diagram illustrating a QuAD system to which the information processing device 10 shown in FIG. 2 belongs. [Figure 4] 3 is a flowchart showing a first operation example of the information processing device 10 shown in FIG. [Figure 5] FIG. 4 is a diagram illustrating a prediction network used in the QuAD system shown in FIG. 3. [Figure 6] FIG. 4 is a diagram illustrating the attention mechanism used in the QuAD system shown in FIG. 3. [Figure 7] 4 is a diagram illustrating estimation of an individual period in the QuAD system shown in FIG. 3. FIG. DETAILED DESCRIPTION OF THE INVENTION

[0019] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0020] (Configuration of information processing device 10) The configuration of an information processing device 10 according to one embodiment of the present invention will be described with reference to FIG.

[0021] The information processing device 10 is, for example, any computer or tablet that belongs to a cloud computing system or other computing system. As will be described in detail later, the information processing device 10 may belong to a computing system called a QuAD system shown in Fig. 3. Here, the QuAD system is designed using two techniques: a convolutional neural network (CNN) used in image or audio processing, and an attention mechanism used in natural language processing.

[0022] The information processing device 10 includes, for example, a control unit 11, a communication unit 12, a storage unit 13, an input unit 14, and an output unit 15.

[0023] The control unit 11 includes a processor such as a CPU (central processing unit) or a GPU (graphics processing unit), a programmable circuit such as an FPGA (field-programmable gate array), a dedicated circuit such as an ASIC (application specific integrated circuit), or any combination thereof. Note that, as will be described in detail later, the control unit 11 executes processes related to the operation of the information processing device 10 while controlling each unit of the information processing device 10.

[0024] The communication unit 12 includes a communication interface. The communication interface may be compatible with, for example, a mobile communication standard such as 4G (4th Generation) or 5G (5th Generation), a wired LAN (Local Area Network) standard, or a wireless LAN standard, but is not limited to these. Note that, as will be described in detail later, the communication unit 12 receives data used in the operation of the information processing device 10 and transmits data obtained by the operation of the information processing device 10.

[0025] The storage unit 13 includes a memory such as a semiconductor memory, a magnetic memory, or an optical memory. The memory may function as, for example, a main storage device, an auxiliary storage device, or a cache memory. Note that, as will be described in detail later, the storage unit 13 stores any data used in the operation of the information processing device 10.

[0026] The input unit 14 includes an input interface. The input interface is, for example, but not limited to, a physical key, a capacitance key, a pointing device, or a touch screen integrated with a display. The input unit 14 accepts an operation to input data used for the operation of the information processing device 10. Note that instead of being provided in the information processing device 10, the input unit 14 may be connected to the information processing device 10 as an external input device.

[0027] The output unit 15 includes an output interface. The output interface is a display such as an LCD (liquid crystal display) or an organic EL (electro luminescence) display, but is not limited to these. The output unit 15 outputs data obtained by the operation of the information processing device 10. Note that the output unit 15 may be connected to the information processing device 10 as an external output device instead of being provided in the information processing device 10.

[0028] (First Operation Example of Information Processing Device 10: QuADNet) A first operation example of the information processing device 10 according to this embodiment will be described with reference to Fig. 4. The first operation example corresponds to the processing executed in QuADNet in the QuAD system shown in Fig. 3.

[0029] [Time series data acquisition processing] Step S100: The control unit 11 of the information processing device 10 receives time series data x indicating a quasi-periodic pattern. t is acquired at a predetermined time period (t=1, ..., T).

[0030] Specifically, the control unit 11 of the information processing device 10 receives time series data x indicating a quasi-periodic pattern from sensors such as d smart meters installed on a production line or the like via the communication unit 12. t Obtain x t is the input matrix X expressed by the following equation (1). twhere t is a natural number from 1 to T and indicates an index of each time point. The control unit 11 of the information processing device 10 then calculates the time series data x t From the above, the input matrix X bundled at w time points is expressed by the following equation (1). t and the output matrix Y at time h, which is expressed by the following equation (2): t Then, the control unit 11 of the information processing device 10 prepares the input matrix X t and the output matrix Y t are stored in the storage unit 13. Note that d, w, and h are natural numbers and can be arbitrarily set via the input unit 14 of the information processing device 10. In addition, the time series data x t is not limited to time series data indicated by sensors such as smart meters installed on production lines, etc., as long as it is time series data indicating a quasi-periodic pattern, but may also be time series data indicated by medical equipment such as electrocardiographs, or other time series data.

[0031]

number

[0032]

number

[0033] In the first operation example, the input matrix X t to the output matrix Y t It is trained to improve the accuracy of multi-temporal prediction.

[0034] [Feature vector generation process] Step S110: The control unit 11 of the information processing device 10 calculates the time series data x acquired in step S100 using a convolutional neural network to which a kernel having multiple channels is applied. t Based on this, a feature vector z t (where t=1, . . . , n) are generated for each of a plurality of time points.

[0035] Specifically, referring also to FIG. 5, the control unit 11 of the information processing device 10 converts the time series data x acquired in step S100 into t The input matrix X prepared from t is regarded as an image having one channel and an image size of w×d. Then, the control unit 11 of the information processing device 10 inputs the input matrix X t As a result, the control unit 11 of the information processing device 10 generates a feature vector z t The control unit 11 of the information processing device 10 can generate the feature vector z t Using the above, a temporal local feature matrix Z expressed by the following equation (3) is constructed. Then, the control unit 11 of the information processing device 10 stores the temporal local feature matrix Z in the storage unit 13. In the following equation (3), T denotes a transposed matrix. In the following equation (3), n (= w - k + 1) is a natural number and indicates the length of the temporal local feature matrix Z in the time axis direction. Furthermore, k is a natural number and indicates the kernel size of the 2D CNN in the time axis direction. Furthermore, c is a natural number and indicates the number of channels of the temporal local feature matrix Z. In the first loop of the learning process, c and k are arbitrarily set via the input unit 14, and in the second and subsequent loops of the learning process, they can be updated by the processing of step S170, which will be described later.

[0036]

number

[0037] Step S120: The control unit 11 of the information processing device 10 passes the time-local feature matrix Z generated in step S110 through a Long Short-Term Memory (LSTM) to generate a feature vector l that takes into account time transitions. t(where t=1, . . . , n) is used to generate a feature matrix L that takes into account time transitions.

[0038] Specifically, referring also to FIG. 5, the control unit 11 of the information processing device 10 inputs the time local feature matrix Z generated in step S110 into a known or arbitrary LSTM capable of expressing time series feature quantities with internal cells, thereby generating a feature vector l taking time transitions into account. t Then, the control unit 11 of the information processing device 10 generates the feature vector l t Using this, a feature matrix L taking into account time transitions is constructed as shown in the following equation (4). Then, the control unit 11 of the information processing device 10 stores the feature matrix L taking into account time transitions in the storage unit 13. Note that in the following equation (4), q is the number of units in the hidden layer of the LSTM, and is a natural number that can be set in advance by the user. q can be set to, for example, 50 or 100, but the present invention is not limited to these.

[0039]

number

[0040] The process of step S120 is optional, and the feature vectors l constituting the feature matrix L generated in step S120 are t Instead of the feature vector z constituting the time local feature matrix Z generated in step S110, t may be used in the process of step S130 described later.

[0041] [Output vector generation process] Step S130: The control unit 11 of the information processing device 10 uses the attention mechanism to calculate the feature vectors l at multiple points in time. t (where t=1, ..., n) and the feature vector l at the earliest point in time among multiple points in time. n Based on the similarity between the feature vector l and t (t=1, ..., n) to generate the output vector a of the attention mechanism.

[0042] Here, the attention mechanism is a method of expressing the output as a weighted sum of n vectors, and outputs an output vector a according to the following equation (5). In the following equation (5), q is a query vector. Also, K is a key matrix (matrix size: q × n). Also, K T is the transposed matrix of the key matrix. The softmax operation is expressed by the following equation (6). ω is a weight vector. V is a value matrix (matrix size: v×n).

[0043]

number

[0044]

number

[0045] Specifically, referring also to FIG. 5, in this example, the control unit 11 of the information processing device 10 calculates the feature vector l generated in step S120 as follows: t Using the feature matrix L composed of the above, in the above equation (5), K = L T and V=L T and q n =l n Then, we apply the attention mechanism shown in Figure 6. This allows us to obtain the feature vector l at multiple points in time. t (where t=1, ..., n) and the feature vector l at the earliest point in time among multiple points in time. n Based on the similarity between the feature vector l and t is weighted to generate an output vector a of the attention mechanism. Then, the control unit 11 of the information processing device 10 stores the output vector a of the attention mechanism in the storage unit 13.

[0046] [Time series data prediction processing] Step S140: The control unit 11 of the information processing device 10 inputs the output vector a of the attention mechanism generated in step S130 to the fully connected layer of the convolutional neural network used in step S110, thereby generating the time series data Y^ after a predetermined period. t Note that Y^ t is a matrix (matrix size: d × h).

[0047] Specifically, with reference to FIG. 5, the control unit 11 of the information processing device 10 executes the following equations (7) and (8). Then, the control unit 11 of the information processing device 10 calculates Y^ t is stored in the storage unit 13. In the following equations (7) and (8), N l is the number of layers, and W i is the weight matrix, and b i is the bias vector, and n i is the number of units in the i-th layer of the fully connected layer (n Nl = dh, n0 = q), and matrix is ​​an operator that reshapes a vector into a matrix (sorts the elements). However, the present invention is not limited to the following equations (7) and (8).

[0048]

number

[0049]

number

[0050] Step S150: The control unit 11 of the information processing device 10 calculates the output matrix Y t and Y^ predicted in step S140 t The error l({Y t}, {Y^ t}).

[0051] Specifically, the control unit 11 of the information processing device 10 calculates the output matrix Y tand Y^ predicted in step S140 t The error l({Y t}, {Y^ t Then, the control unit 11 of the information processing device 10 calculates the error l({Y t}, {Y^ t}) is stored in the storage unit 13. Note that, as a method for calculating the error, it is preferable to use an L1 loss function that is highly robust against outliers, but the present invention is not limited to this.

[0052]

number

[0053] Step S160: The control unit 11 of the information processing device 10 calculates the error l({Y t}, {Y^ t}) is below a predetermined threshold. t}, {Y^ t If the error l({Y t}, {Y^ t If {}) is not less than or equal to the predetermined threshold, the process proceeds to step S170.

[0054] Here, the predetermined threshold value can be arbitrarily set via the input unit 14 of the information processing device 10, taking into consideration the accuracy of multi-point prediction. However, the learning termination criterion in the present invention is t}, {Y^ t The determination is not limited to whether or not the parameter values ​​are equal to or less than a predetermined threshold. For example, when the number of learning loops reaches a preset value, the control unit 11 of the information processing device 10 may determine that learning has ended and terminate updating of the values ​​of various parameters.

[0055] Step S170: The control unit 11 of the information processing device 10 updates the values ​​of various parameters that configure the convolutional neural network used in step S110 and the LSTM used in step S120.

[0056] Specifically, the control unit 11 of the information processing device 10 updates the various values ​​of the 2D CNN used in step S110, the LSTM used in step S120, and the weight matrix and bias vector included in the fully connected layer using a known or arbitrary optimization method such as stochastic gradient descent (SGD) or Adam. Then, the control unit 11 of the information processing device 10 stores the various updated values ​​in the storage unit 13. As a result, the calculations of the 2D CNN and the LSTM in the second and subsequent loops of the learning process are performed using the various updated values. Then, the process returns to step S110.

[0057] According to the first operation example, it is possible to provide a technique for predicting time series data showing a quasi-periodic pattern, which enables estimation of individual periods, which will be described later in the second operation example, and anomaly detection, which will be described later in the third to fifth operation examples. Note that the control unit 11 of the information processing device 10 may simultaneously perform anomaly detection in the second to fifth operation examples, which will be described later, in real time.

[0058] In the second to fifth operation examples below, a prediction model constructed using the various parameters optimized in the first operation example is used.

[0059] (Second Operation Example of Information Processing Device 10: Estimation of Individual Periods) A second operation example of the information processing device 10 according to this embodiment will be described with reference to Fig. 7. The second operation example corresponds to estimation of an individual period in the QuAD system shown in Fig. 3.

[0060] In the second operation example, the control unit 11 of the information processing device 10 calculates the time series data x based on a time point, among the plurality of time points, excluding the earliest time point (n=w−k+1), and based on the time point at which the value of the element of the weight vector corresponding to the similarity calculated in step S130 in the first operation example is maximum. t The individual periods at each time are estimated.

[0061] Specifically, the control unit 11 of the information processing device 10 inputs each element of the weight vector ω calculated in step S130 in the first operation example and expressed by the following equation (10) into the following equation (11), thereby calculating the time series data x t In the example shown in FIG. 7, the individual period is approximately 91 cycles. In the following equation (11), the argmax function is in the range of 1≦s≦(w−k+1)−s c This function returns the value (time) of s at which the value of the element of the weight vector ω is maximum. Here, the upper limit of s is s=(w-k+1)-s c The reason for the restriction is that, as shown in Figure 7, the elements of the weight vector ω close to s=w-k+1, which corresponds to the current time, have a high similarity regardless of the quasi-periodicity, and therefore must be removed in order to correctly estimate the quasi-periodic pattern. c indicates the disconnection time, and can be appropriately set via the input unit 14 of the information processing device 10 as the time at which the partial correlation coefficient decreases.

[0062]

number

[0063]

number

[0064] Additionally, the control unit 11 of the information processing device 10 calculates the estimated value c of the individual period calculated by the above formula (11). t Using the above, the robust estimator C tIn the following equation (12), the surrounding n e The period is added to the count of each period. p represents an estimate of the globally fundamental base period, and is estimated by the following equation (13).

[0065]

number

[0066]

number

[0067] Note that a change in the lot size of a production machine or the like appears as a change in the basic period. Therefore, the control unit 11 of the information processing device 10 calculates the estimated value C of the individual period estimated by the above equation (11). t is the estimated fundamental period b estimated by the above equation (13). p If it is determined that the deviation from the target value by the same amount is constant, a signal to change the lot may be sent to the system, and the learning process in the first example of operation may be performed again. When the learning process is performed again, the previous learning results may be treated as a pretrained network and fine-tuned, or the entire network may be retrained. In either case, all parameters will be changed.

[0068] According to the second operation example, it is possible to estimate individual periods at each time of time series data that exhibits a quasi-periodic pattern.

[0069] (Third Operation Example of Information Processing Device 10: QuADNet-P) A third operation example of the information processing device 10 according to this embodiment will be described. The third operation example corresponds to the processing executed by QuADNet-P in the QuAD system shown in FIG.

[0070] In the third operation example, the control unit 11 of the information processing device 10 determines whether or not there is an abnormality in the acquired time series data based on a comparison between the time series data acquired in the same manner as in step S100 in the first operation example and the time series data predicted in the same manner as in step S140 in the first operation example.

[0071] Specifically, the control unit 11 of the information processing device 10 performs the same process as in step S100 in the first operation example to generate time series data x t From the output matrix Y at time h t where Y t The elements of (Y t ) is Similarly to step S100 in the first operation example, the control unit 11 of the information processing device 10 generates time series data x t From the input matrix X bundled at w time points t Then, the control unit 11 of the information processing device 10 inputs the input matrix X t By inputting the above, the multi-point forecast value Y^ t where Y^ t The elements of (Y^ t ) is =(QuADNet(X t )) is Here, i (where i=1, ..., d) is an index for identifying a sensor such as a smart meter. Also, s (where s=1, ..., h) is an index corresponding to each time point. Then, the control unit 11 of the information processing device 10 calculates (Y t ) is and (QuADNet(X t )) is By inputting the above into the following equation (14), the mean absolute error (MAE) is calculated as MAE(X t ,Y t Then, the control unit 11 of the information processing device 10 calculates the MAE(X t ,Y t) is equal to or greater than a predetermined threshold, it is determined that an abnormality exists. t ,Y t If (Y ) is less than a predetermined threshold, it is determined that there is no abnormality. Then, the control unit 11 of the information processing device 10 outputs the result of the determination via the output unit 15. t ) is and (QuADNet(X t )) is It is preferable to use MAE as an evaluation index for the error from the viewpoint of robustness against spike data, but the present invention is not limited to this. In addition, the predetermined threshold can be set appropriately via the input unit 14.

[0072]

number

[0073] According to the third operation example, an abnormality can be detected in time-series data that exhibits a quasi-periodic pattern.

[0074] (Fourth Operation Example of Information Processing Device 10: QuADNet-D) A fourth operation example of the information processing device 10 according to this embodiment will be described. The fourth operation example corresponds to the processing executed by QuADNet-D in the QuAD system shown in FIG.

[0075] In the fourth operation example, the control unit 11 of the information processing device 10 determines whether or not there is an abnormality in the time series data acquired in the same manner as in step S100 in the first operation example, based on a probability distribution constructed using the elements of the weight vector ω calculated in step S130 in the first operation example.

[0076] Specifically, the control unit 11 of the information processing device 10 uses the elements of the weight vector ω calculated in step S130 in the first operation example to construct a Dirichlet distribution f(ω,α) expressed by the following equation (15): sis the set of weight vectors {ω t}, and is appropriately obtained by maximum likelihood estimation from {\displaystyle \mathbb {\displaystyle \mathbb {f(ω,α)}}. That is, by performing maximum likelihood estimation, a Dirichlet distribution f(ω,α) is obtained as the distribution of the weight vector ω, and the presence or absence of an abnormality is determined by evaluating the degree of fit to the maximum likelihood estimated Dirichlet distribution. Then, the control unit 11 of the information processing device 10 outputs the result of this determination via the output unit 15. Note that, since the weight vector ω can be interpreted as a probability vector whose sum is 1, it is preferable to use a Dirichlet distribution, which is a distribution for probability vectors whose sum is 1, for the probability distribution in the fourth operation example, but the present invention is not limited to this.

[0077]

number

[0078] According to the fourth operation example, an abnormality can be detected in time-series data that exhibits a quasi-periodic pattern.

[0079] (Fifth Operation Example of Information Processing Device 10: QuADNet-C) A fifth operation example of the information processing device 10 according to this embodiment will be described. The fifth operation example corresponds to the processing executed by QuADNet-C in the QuAD system shown in FIG.

[0080] In the fifth operation example, the control unit 11 of the information processing device 10 determines whether or not there is an abnormality in the time series data acquired in the same manner as step S100 in the first operation example by constructing an empirical distribution from the frequency of the individual period estimate {Ct} obtained in the second operation example.

[0081] Specifically, the control unit 11 of the information processing device 10 inputs the individual period estimate {Ct} obtained in the second operation example into the empirical distribution expressed by the following equation (16) to determine whether or not an abnormality exists. That is, P in the following equation (16) is also referred to as an empirical likelihood function, and when a new {Ct} is obtained, by substituting it for P, a quantity corresponding to the degree of fit, which serves as an index for determining whether or not an abnormality exists, is obtained. Note that in the following equation (16), I is an indicator function that gives 1 if true and 0 if false. Also, ε (>0) is a small parameter that can be set as appropriate to prevent divergence when taking the logarithm of P(C=i).

[0082]

number

[0083] According to the fifth operation example, an abnormality can be detected in time-series data that exhibits a quasi-periodic pattern.

[0084] In the third to fifth operation examples described above, the presence or absence of an abnormality is determined based on likelihood, but the control unit 11 of the information processing device 10 may determine the presence or absence of an abnormality after performing post-processing based on the following steps 1 to 5 in order to identify robust abnormal sections.

[0085] Step 1: Log likelihood is calculated by p Average the values. Step 2: Determine whether abnormal or normal based on the threshold. Step 3: Expand and / or contract the abnormal section by γ1. Step 4: The length of the abnormal section is a l If it is below this, it is judged as quasi-abnormal, and a l If it is equal to or greater than this, it is determined to be abnormal. Step 5: Set the abnormal section as γ2>max{γ1,a l} by .

[0086] Step 1 above corresponds to a smoothing process, and is a process for correcting deviations in likelihood within one period. Steps 3 and 5 correspond to closing in image processing, and are processes for integrating sections so that when there is only one or two intervals between abnormal sections, they can be considered to be the same abnormal section. Step 4 corresponds to opening (contraction and / or expansion) in image processing, and is a process for removing short-term abnormalities that are less than half a period. Note that b p may be the same as the above equations (12) and (13). In addition, the log-likelihood refers to the logarithm of the negative MAE in QuADNet-P, the logarithm of f in QuADNet-D, and the logarithm of P in QuADNet-C.

[0087] While the present invention has been described with reference to the drawings and examples, it should be noted that those skilled in the art may make various modifications and alterations based on this disclosure. Therefore, it should be noted that these modifications and alterations are within the scope of the present invention. For example, functions included in each unit or step may be rearranged so as not to cause logical inconsistencies, and multiple components or steps may be combined or divided into one. Furthermore, an embodiment in which the configuration and operation of the information processing device 10 are distributed across multiple computers that can communicate with each other is also possible.

[0088] As a modified example, an embodiment is also possible in which a general-purpose computer functions as the information processing device 10 according to the above-described embodiment. Specifically, a program describing the processing content for realizing each function of the information processing device 10 according to the above-described embodiment is stored in the memory of the general-purpose computer, and the program is read and executed by a processor. Therefore, the present invention can also be realized as a program executable by a processor or a non-transitory computer-readable medium storing the program. [Explanation of symbols]

[0089] 10. Information processing equipment 11 Control section 12 Communications Department 13 Storage section 14 Input section 15 Output section

Claims

1. An information processing device including a control unit, The control unit a time series data acquisition process for acquiring time series data showing a quasi-periodic pattern over a predetermined period; a feature vector generation process for generating, for each of a plurality of time points, a feature vector having a feature amount for each of the channels as an element based on the acquired time series data, using a convolutional neural network to which a kernel having a plurality of channels is applied; an output vector generation process that uses an attention mechanism to weight the feature vectors at the plurality of time points based on a similarity between the feature vectors at the plurality of time points and the feature vector at the earliest time point among the plurality of time points, thereby generating an output vector from the attention mechanism; a time series data prediction process for predicting the time series data after the predetermined period by inputting the output vector into a fully connected layer of the convolutional neural network; An information processing device that executes the above.

2. 2. The information processing device according to claim 1, The control unit estimates individual periods at each time point of the acquired time series data based on a time point among the plurality of time points, excluding the earliest time point, at which a value of an element of a weight vector corresponding to the similarity is maximum.

3. 3. The information processing device according to claim 1, The control unit determines whether or not there is an abnormality in the acquired time series data based on a comparison between the acquired time series data and the predicted time series data.

4. 3. The information processing device according to claim 1, The control unit determines whether or not there is an abnormality in the acquired time-series data based on a probability distribution configured using elements of a weight vector corresponding to the similarity.

5. 3. The information processing device according to claim 2, The control unit determines whether or not there is an abnormality in the acquired time-series data based on an empirical distribution constructed using the estimated frequencies of the individual periods.

6. An information processing method executed by an information processing device, a time-series data acquisition step of acquiring time-series data showing a quasi-periodic pattern for a predetermined period; a feature vector generation step of generating, for each of a plurality of time points, a feature vector having a feature amount for each of the channels as an element based on the acquired time series data by using a convolutional neural network to which a kernel having a plurality of channels is applied; an output vector generation step of generating an output vector of the attention mechanism by weighting the feature vectors at the plurality of time points based on a similarity between the feature vectors at the plurality of time points and the feature vector at the earliest time point among the plurality of time points; a time series data prediction step of predicting the time series data after the predetermined period by inputting the output vector into a fully connected layer of the convolutional neural network; An information processing method, including:

7. On the computer, a time-series data acquisition step of acquiring time-series data showing a quasi-periodic pattern for a predetermined period; a feature vector generation step of generating, for each of a plurality of time points, a feature vector having a feature amount for each of the channels as an element based on the acquired time series data by using a convolutional neural network to which a kernel having a plurality of channels is applied; an output vector generation step of generating an output vector of the attention mechanism by weighting the feature vectors at the plurality of time points based on a similarity between the feature vectors at the plurality of time points and the feature vector at the earliest time point among the plurality of time points; a time series data prediction step of predicting the time series data after the predetermined period by inputting the output vector into a fully connected layer of the convolutional neural network; A program that executes.

Images