WO2021149528A1 - イベント予測システム、イベント予測方法およびプログラム - Google Patents

イベント予測システム、イベント予測方法およびプログラム Download PDF

Info

Publication number
WO2021149528A1
WO2021149528A1 PCT/JP2021/000606 JP2021000606W WO2021149528A1 WO 2021149528 A1 WO2021149528 A1 WO 2021149528A1 JP 2021000606 W JP2021000606 W JP 2021000606W WO 2021149528 A1 WO2021149528 A1 WO 2021149528A1
Authority
WO
WIPO (PCT)
Prior art keywords
time
prediction
sensor data
event
series
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/JP2021/000606
Other languages
English (en)
French (fr)
Japanese (ja)
Inventor
崇人 本田
靖子 櫻井
光希 川畑
保志 櫻井
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Osaka NUC
Original Assignee
Osaka University NUC
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Osaka University NUC filed Critical Osaka University NUC
Priority to US17/793,388 priority Critical patent/US20230058585A1/en
Priority to JP2021573071A priority patent/JP7440938B2/ja
Publication of WO2021149528A1 publication Critical patent/WO2021149528A1/ja
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B23/00—Testing or monitoring of control systems or parts thereof
    • G05B23/02—Electric testing or monitoring
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0201—Market modelling; Market analysis; Collecting market data
    • G06Q30/0202—Market predictions or forecasting for commercial activities
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N7/00—Computing arrangements based on specific mathematical models
    • G06N7/01—Probabilistic graphical models, e.g. probabilistic networks
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30—Computing systems specially adapted for manufacturing

Definitions

  • the present invention relates to an event prediction technique based on time series sensor data.
  • Non-Patent Documents 2, 17, 19, 22, 24, 25 are typical technologies, and there are many methods for analyzing and predicting sensor data based on these (Non-Patent Document 13). ..
  • Non-Patent Document 15 has the ability to estimate a nonlinear dynamic system in real time from a large amount of multidimensional sensor data that continues to be generated, and to continue to predict the future adaptively.
  • this method uses a sensor stream as an input and shows high performance in predicting the measured value of the sensor data, it does not correspond to the prediction of event data such as normal / abnormal.
  • Non-Patent Documents 8, 10, 11, 16, 28, 29, 31 Matsubara et al.
  • Non-Patent Document 18 proposed TriMine as an analysis method for large-scale event tensors. TriMine classifies given data into multiple topics and detects potential trends and patterns, but targets discrete event data such as click logs on the web and time series such as IOT sensor data. It is not possible to express the dynamic pattern of a sequence or its group (regime), and the problems to be dealt with are different. In addition, TriMine does not have the ability to predict events.
  • Non-Patent Documents 3, 9, 26, 27 Research on analysis of nonlinear dynamic characteristics based on Deep Neural Network is also active (Non-Patent Documents 3, 9, 26, 27).
  • Qin et al. Proposed a method for predicting stock prices with high accuracy by modeling important dimensions in the input time series and important dimensions in the special space after dimension reduction over two layers. ..
  • the method of modeling the event occurrence intensity is the mainstream (Non-Patent Documents 5, 6, 20, 30).
  • RMTPP Non-Patent Document 5 proposes a non-linear model for predicting the time and type of the next event to occur from the past event history.
  • these methods target categorical data composed only of event history, and cannot predict events by continuous data composed of measured values from sensors.
  • the neural hawkes process A neutrally self-modulating multivariate point process.
  • NIPS pages 6757-6767, 2017. Y. Qin, D. Song, H. Chen, W. Cheng, G. Jiang, and G. W. Cottonrell.
  • IJCAI pages 2627-2633, 2017. T.Rakthanmanon, B.J.L.Campana, A.Mueen, G.E.A.P.A.Batista, M.B.Westover, Q.Zhu, J.Zakaria, and E.J.Keogh.Searching and mining trillions of time series subsequences under dynamic time warning.
  • the present invention has been made in view of the above, and provides an event prediction system, a method and a program for time-series tensor data, which enables long-term and highly accurate event prediction through data summarization processing. It is to provide.
  • the first feature amount extraction unit of the computer is a time series in which the first feature amount extraction unit of the computer is continuously collected from a plurality of types of sensors arranged in a plurality of observation targets and stored in the storage unit.
  • the model parameters of the multifaceted dynamic pattern are continuously extracted from the sensor data and stored in the storage unit, and the second feature quantity extraction unit of the computer stores the model parameters and the time series sensor data in the storage unit.
  • the time-series sensor data is read from the storage unit, sequentially featured into summary information including modeling information and its error information, and stored in the storage unit, and the prediction unit of the computer stores the summary information. It is read from the unit and used as an input, and the occurrence probability of a predetermined event at a predetermined time ahead is output.
  • the program according to the present invention continuously extracts model parameters of a multifaceted dynamic pattern from time-series sensor data continuously collected from a plurality of types of sensors arranged in a plurality of observation targets.
  • the first feature amount extracting means, the second feature amount extracting means for sequentially characterizing the time-series sensor data into summary information including modeling information and its error information using the model parameters, and the summary information.
  • the computer functions as a predictive means for outputting the occurrence probability of a predetermined event at a predetermined time ahead as input.
  • time-series sensor data is continuously collected from a plurality of types of sensors arranged in a plurality of observation targets, and model parameters of a multifaceted dynamic pattern are collected from the collected time-series sensor data.
  • the extraction is continuously performed by the first feature amount extraction means.
  • the second feature amount extracting means sequentially features the time-series sensor data into summary information including modeling information and error information thereof using the model parameters.
  • the prediction means outputs the probability of occurrence of a predetermined event at a predetermined time ahead by inputting the summary information.
  • the change points and potential behaviors of the patterns can be determined from, for example, time transitions and multifaceted viewpoints between observation targets. Be grasped.
  • the sensor may be arranged directly on the observation target or in a manner in which the observation target can be observed remotely.
  • FIG. 1 It is an overall block diagram which shows one Embodiment of the event prediction system which concerns on this invention. It is a figure which shows an example of the processing state of the information taken in from the smart factory data which this invention is an application example, (a) shows the original sensor data, (b) shows the pattern detection result from the original data. , (C), (d) are diagrams showing typical regime examples of a case where an emergency stop occurs after a predetermined time (d) and a case where the emergency stop is not performed (c) based on the original data. It is a figure which shows the outline of the proposed model which concerns on this invention. It is a transition diagram for demonstrating the basic concept of the proposed algorithm which concerns on this invention.
  • the present invention preferably relates to an event prediction method for large-scale time series sensor data.
  • the present invention comprehensively analyzes and summarizes a multifaceted time series pattern based on a plurality of viewpoints from, for example, factory equipment sensor data composed of a triplet of (facility, sensor, time), and in the future.
  • technology for long-term event prediction More specifically, when time-series data consisting of measured values of sensor data such as rotation speed, operating voltage, and equipment temperature in each equipment installed in the factory is given, (a) basic time-series. By extracting patterns, common patterns between each facility and patterns unique to each facility, and statistically summarizing them, (b) predict future events. Moreover, these processes are (c) linear with respect to the data size.
  • this prediction system will occur in the future by grasping the number of typical patterns (hereinafter referred to as regimes) included in time series data and the points of change from various angles and accurately grasping the operating status of the system. Predict events. More specifically, when large-scale time-series sensor data collected from a plurality of sensors is given in a plurality of facilities, an event after a predetermined time, that is, one l s step ahead is predicted.
  • FIG. 1 shows an overall block diagram of an event prediction system (hereinafter, prediction system 1) according to the present invention.
  • This prediction system 1 transmits large-scale time-series sensor data related to the operating status from each of the sensor groups 21 installed in the observation targets 20, ...
  • a computer having a control unit 10 including a processor (CPU) that extracts a feature amount from each of the captured time-series data and executes an event prediction process after a predetermined time is provided. ..
  • machine learning is used, and the parameters applied to the prediction processing are updated through machine learning. Details of FIG. 1 will be described later.
  • FIG. 2 is sensor data from a smart factory as an example of the observation target 20 (FIG. 1), and shows information to be used (input) for prediction processing.
  • FIG. 2A shows the original sensor data, and three sensor values (rotational speeds) collected as an example of each sensor group 21 (FIG. 1) from the five equipments (# 1 to # 5). : Speed, operating voltage: Load, equipment temperature: Temp).
  • Speed operating voltage: Load
  • equipment temperature Temp
  • FIG. 2A the part painted with a black rectangle indicates that the corresponding equipment is in an emergency stop.
  • the waveform of the operating voltage: Load in FIG. 2A generally overlaps with the waveform of the rotation speed: Speed.
  • the prediction system 1 analyzes time-series data obtained from a plurality of facilities at the same time to generate a multifaceted pattern, that is, not only a time transition of a pattern in each facility but also a pattern common or different among the facilities. It is possible to detect.
  • the left side of FIGS. 2 (c) and 2 (d) shows the segmentation result.
  • ⁇ 1 to ⁇ 5 on the right side represent common time-series patterns (that is, regimes), and visualize the transitions between them.
  • the value of p200 is the emergency stop probability at 200 steps ahead output by this prediction system when the partial sequence corresponding to the left side figure of FIGS. 2C and 2d and the pattern detection result thereof are given.
  • thick arrows are displayed between the regimes in which more transitions are detected.
  • the size of the circle indicates the size of the period during which the regime occurs.
  • the rotation speed Speed increases ( ⁇ 5 ) before the equipment stops in an emergency, and this tendency is expressed by the appearance of transitions between the regimes ⁇ 4 and ⁇ 5.
  • the prediction system 1 accurately predicts an emergency stop, and p200 shows a high value. That is, by detecting the potential pattern contained in the data, not only the process leading to the emergency stop can be analyzed from various angles, but also long-term and highly accurate prediction can be made by using the summary information.
  • transitions without signs of an emergency stop such as regimes ⁇ 2 , ⁇ 3 , ⁇ 2 , ⁇ 1 , and ⁇ 2 are observed, and p200 also shows a low value.
  • This prediction system 1 predicts the equipment alert of the l s step ahead from the given time-series tensor X, and the processing required for that is shown below.
  • the alert label y (t e + l s ) at the l s step destination is predicted based on the following equation (1).
  • t s : t e represents the window of the sequence used for prediction (a predetermined period from the present time to the past direction), and F is a proposed model.
  • the prediction system 1 executes the following three processes (P1), (P2), and (P3).
  • a storage unit 100 for example, a display unit 121 for displaying a window described later, and an operation unit 122 for receiving an instruction from the outside are connected to the control unit 10.
  • the storage unit 100 includes a control program storage unit 101, a data stream storage unit 102 that stores time-series sensor data input from each sensor group 21, and a neural network that constitutes artificial intelligence (AI) applied to prediction processing.
  • a parameter storage unit 103 that stores model parameters (weights of each edge, etc.) is provided.
  • the control program storage unit 101 stores program data for executing the event prediction process described later and various necessary arithmetic expression data.
  • the storage unit 100 also performs processing "(P1) multifaceted detection of potential dynamic patterns", “(P2) feature extraction based on dynamic patterns” and processing described later. It has a work area (storage unit) that temporarily stores each data obtained during the execution of "(P3) l s step ahead long-term prediction”.
  • the control unit 10 functions as a data acquisition processing unit 11, a feature amount extraction unit 12, a prediction unit 13, and a parameter update unit 14 when the control program is executed.
  • the data acquisition processing unit 11 acquires time-series sensor data from the sensor group 21 of each observation target 20 (each facility in the factory) via the network 110.
  • the feature extraction unit 12 executes the processes “(P1) multifaceted detection of potential dynamic patterns” and “(P2) feature extraction based on dynamic patterns”, which will be described later.
  • the prediction unit 13 executes the process “(P3) l s step ahead long-term prediction”. In the present embodiment, the prediction unit 13 applies the parameters from the parameter storage unit 103 to perform the prediction process. Details of each process will be described later.
  • the machine learning device 30 includes a control unit 300 and a storage unit 310 composed of a computer having a built-in processor, and also includes a display unit 321 and an operation unit 322.
  • the storage unit 310 includes a learning program storage unit 311, a data stream storage unit 312, and a parameter storage unit 313.
  • the data stream storage unit 312 captures the time-series sensor data input from each sensor group 21 by communication or via an external memory, or fetches and stores the data once written in the data stream storage unit 102.
  • the control unit 300 functions as a data acquisition processing unit 301, a feature amount extraction unit 302, and a machine learning unit 303 by executing a learning program from the learning program storage unit 311.
  • the data acquisition processing unit 301 is the same as the data acquisition processing unit 11, and the acquisition period of the acquired data can be appropriately set automatically or manually (for example, for the latest one week).
  • the feature amount extraction unit 302 is provided as necessary, and confirms the processing by appropriately adjusting the conditions of the processing (P1) and (P2) according to, for example, a change in factory equipment or other changes in the situation.
  • the machine learning unit 303 preferably performs machine learning by applying, for example, "supervised learning” to the time-series sensor data for the latest predetermined period, and stores the parameters that are the learning results in the parameter storage unit 313. Then, if necessary, the parameter storage unit 103 is updated via the parameter update unit 14 or in response to an instruction from the operation unit 322 of the machine learning device 30.
  • various modes can be adopted in addition to the mode of the separate machine learning device 30.
  • the input data may be taken out from the data stream storage unit 102 for a predetermined period.
  • the learning may be executed by using the prediction unit 13 by using the system stop period (for example, at night), and the parameter which is the learning result may be updated.
  • the multidimensional time series tensor X can be expressed as ⁇ m, r, S, ⁇ , F ⁇ with m segments and r regimes.
  • this prediction system statistically models the multidimensional time series tensor X based on the obtained regime information and extracts important features.
  • HMM hidden Markov model
  • k indicates the number of latent states of HMM.
  • the output probability B is generated from the multidimensional Gaussian distribution.
  • ⁇ ) of ⁇ is as shown in the following equation (Equation 1). It is calculated.
  • Cost M (M) indicates the model cost for expressing the model M
  • M) indicates the coding cost of the tensor X given the model M.
  • the coding cost of X when a model parameter is given can be expressed as follows (Equation 6) using negative log-likelihood by information compression using Huffman coding. ..
  • ⁇ SPLITCAST (P3): Among ⁇ Z, ⁇ , a partial sequence of a window t s : t e ⁇ Z (t s : t e ), ⁇ (t s : t e ) ⁇ is a feature that predicts a failure. Is extracted, and the failure label y (t e + l s ) at the l s destination is predicted.
  • the fundamental question in time series analysis is whether there are hidden structures inherent in the time series data.
  • the multidimensional time series tensor X dealt with here has features from a plurality of viewpoints. That is, the characteristics of the time domain and the characteristics of the equipment domain. Specifically, the time-series sensor data obtained from the smart factory has a time transition pattern of each process and a pattern peculiar to the equipment. Therefore, in the following, we will simultaneously perform multifaceted pattern discovery and grouping, which briefly summarizes the underlying structure of a given time-series tensor.
  • the V-Assignment can detect the change point of the X pattern based on the model parameters of the regime (steps 5 to 7 in Table 2).
  • the transition diagram of FIG. 4 is shown. Estimate the pattern transition between given regimes by connecting the transitions of the two regimes ⁇ 1 , ⁇ 2 ⁇ and comparing the coding costs of the two regimes at each time.
  • This algorithm calculates the coding cost Cost T (X
  • ⁇ ) -ln P (X
  • H-Assignment effectively extracts equipment-specific patterns. Specifically, when the tensor X and the model parameters ⁇ 1 , ⁇ 2 ⁇ are given, the algorithm 2 calculates the coding cost when the segment of the equipment i is assigned to a certain regime ⁇ as follows (Equation 10). And assign the segment of equipment i to the regime where the cost is lower.
  • z i (t) indicates the set ⁇ , ⁇ of the mean and variance of all data points belonging to the same state as itself. That is, the dimension of the latent state tensor is Z ⁇ R w ⁇ 2d ⁇ n .
  • the coding error of the measured value x ij (t) ⁇ X of the sensor j of the equipment i at time t is expressed by the posterior probability p (x ij (t)
  • this prediction system combines a feature extraction method based on a probabilistic model and a deep learning method to learn characteristic time-series patterns extracted from actual data, so that it can be learned with a smaller network and solve the problem of overfitting. Achieve efficient and effective alert label prediction while mitigating.
  • LSTM Long-short term memory
  • Non-Patent Document 9 is one of the deep learning models that treats input samples as time series data and enables learning of high-dimensional nonlinear dynamics.
  • LSTM replaces the unit in the middle layer of RNN (Recurrent neural network) with a special structure called a memory unit. It uses three types of input gate, output gate, and forgetting gate, and the unit value c t at time t. And the output value h t of the unit.
  • Each i t the output values of each gate, o t, When f t, forward propagation of LSTM is expressed by the following equation (11).
  • the sigmoid function is used as the activation function.
  • LSTM can learn the long-term dependence of the input sequence given by the memory unit, so it remembers the features that are particularly important for equipment failure in the process of regime transition and state transition inside the regime. At the same time, it is considered that a feature vector summarizing the latest operating status of the equipment is extracted.
  • Equation 13 The objective function that the model in this prediction system should minimize is BCE (Binary cross entropy). If the batch size at the time of model learning is N and the output value in this prediction system for each input sample i is y ⁇ i , It is represented as shown by (Equation 13).
  • V-Assignment, H-Assignment, and ModelEstimation require O (wdnk 2 ) complexity to estimate the coding cost and model parameters.
  • w is the number of equipment
  • d is the number of dimensions
  • n is the length of the time series
  • the number of iterations #iter and the number of hidden states k are very small constants and can be ignored. Therefore, the computational complexity of RegimeGeneration is O (wdn).
  • the amount of calculation is O (wdn) because it outputs the error when modeling with the latent state of each equipment, each sensor, and each time.
  • O (wdn) because it outputs the error when modeling with the latent state of each equipment, each sensor, and each time.
  • LR Logistic regression
  • RNN Recurrent neural network
  • GRU Gate recurrent unit
  • LSTM LSTM
  • the dataset used was Speed, which was installed in 55 factory equipment that had been in operation for three months from October 2017 at Mitsubishi Heavy Industries Engine & Turbo Charger Co., Ltd. and was processing bearings and housings. , Operating voltage (Load), and equipment temperature (Temp), which are acquired at 5-second intervals.
  • the learning sample is generated in the sliding window, and the sample when the equipment itself is not in operation is omitted. Since the number of samples during normal operation is 62983 and the number of samples before emergency stop is 1069, which causes a bias in learning, the number of samples during normal operation is aligned with the number of samples during emergency stop, and as a result, 1069 x 2 samples are used. An experiment was conducted.
  • FIG. 5 is a comparison diagram of the accuracy when the number of predicted destination steps l s is changed.
  • the type notation of the comparative example and the data display order (left and right) correspond.
  • FIG. 9 is a diagram showing the relationship between the number of training samples and the prediction accuracy. This prediction system shows higher performance than the comparative example even with a small number of samples, and as the number of training samples increases, it is possible to predict a failure event with higher accuracy.
  • this prediction system conducts experiments using, for example, actual data obtained from factory equipment, and this prediction system appropriately models complex time-series patterns to make long-term failure prediction highly accurate. It was confirmed that it could be done, and that it achieved a significant improvement in accuracy and performance compared to the existing comparative examples.
  • the event prediction system has a multi-faceted dynamic pattern of model parameters from time-series sensor data continuously collected from a plurality of types of sensors arranged in a plurality of observation targets.
  • a first feature amount extraction means that continuously performs extraction, and a second feature amount extraction that sequentially features the time series sensor data into summary information including modeling information and its error information using the model parameters. It is preferable to include means and predictive means for outputting the occurrence probability of a predetermined event at a predetermined time ahead by inputting the summary information.
  • the first feature amount extraction unit of the computer is a time series in which the first feature amount extraction unit of the computer is continuously collected from a plurality of types of sensors arranged in a plurality of observation targets and stored in the storage unit.
  • the model parameters of the multifaceted dynamic pattern are continuously extracted from the sensor data and stored in the storage unit, and the second feature quantity extraction unit of the computer stores the model parameters and the time series sensor data in the storage unit.
  • the time-series sensor data is read from the storage unit, sequentially featured into summary information including modeling information and its error information, and stored in the storage unit, and the prediction unit of the computer stores the summary information. It is preferable to read from the unit and use it as an input to output the occurrence probability of a predetermined event at a predetermined time ahead.
  • the program according to the present invention continuously extracts model parameters of a multifaceted dynamic pattern from time-series sensor data continuously collected from a plurality of types of sensors arranged in a plurality of observation targets.
  • the first feature amount extracting means, the second feature amount extracting means for sequentially characterizing the time-series sensor data into summary information including modeling information and its error information using the model parameters, and the summary information. It is preferable to make a computer function as a predictive means for outputting the occurrence probability of a predetermined event at a predetermined time ahead as input.
  • time-series sensor data is continuously collected from a plurality of types of sensors arranged in a plurality of observation targets, and model parameters of a multifaceted dynamic pattern are collected from the collected time-series sensor data.
  • the extraction is continuously performed by the first feature amount extraction means.
  • the second feature amount extracting means sequentially features the time-series sensor data into summary information including modeling information and error information thereof using the model parameters.
  • the prediction means outputs the probability of occurrence of a predetermined event at a predetermined time ahead by inputting the summary information.
  • the change points and potential behaviors of the patterns can be determined from, for example, time transitions and multifaceted viewpoints between observation targets. Be grasped.
  • the sensor may be arranged directly on the observation target or in a manner in which the observation target can be observed remotely.
  • the first feature amount extracting means detects the dynamic pattern by performing a segment and its patterning in the time direction and between the observation targets. According to this configuration, since the dynamic pattern is extracted from various angles, it is possible to reduce the amount of data required for processing while suppressing a decrease in accuracy.
  • the first feature amount extracting means sets the number of the segments by using a cost function. According to this configuration, in the segmentation of the time series sensor data, the number of segments is set to the optimum value in consideration of the data amount and the processing time by the cost function.
  • the prediction means obtains the probability of occurrence of the predetermined event based on the parameters set in the neural network model. According to this configuration, highly accurate prediction is possible with a model having a small and simple structure.
  • the prediction means applies LSTM (Long-short term memory) to the neural network model.
  • LSTM Long-short term memory
  • the LSTM can be applied to a deep learning model, and the long-term dependence of the input sequence can be learned, so that it is possible to make a highly accurate prediction for a long period of time.
  • the summary information obtained by the second feature amount extracting means is taken in for a predetermined period, machine learning is performed by a learning prediction means having the same configuration as the prediction means, and the learning result is obtained. It is preferable to provide a machine learning device that updates the parameters to the prediction means. According to this configuration, it is possible to gradually improve the prediction accuracy.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Business, Economics & Management (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Strategic Management (AREA)
  • Finance (AREA)
  • Accounting & Taxation (AREA)
  • Development Economics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Algebra (AREA)
  • General Business, Economics & Management (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Game Theory and Decision Science (AREA)
  • Computational Mathematics (AREA)
  • Economics (AREA)
  • Mathematical Analysis (AREA)
  • Marketing (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Automation & Control Theory (AREA)
  • Testing And Monitoring For Control Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
PCT/JP2021/000606 2020-01-22 2021-01-12 イベント予測システム、イベント予測方法およびプログラム Ceased WO2021149528A1 (ja)

Priority Applications (2)

Application Number Priority Date Filing Date Title
US17/793,388 US20230058585A1 (en) 2020-01-22 2021-01-12 Event forecasting system, event forecasting method, and storage medium
JP2021573071A JP7440938B2 (ja) 2020-01-22 2021-01-12 イベント予測システム、イベント予測方法およびプログラム

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP2020008388 2020-01-22
JP2020-008388 2020-01-22

Publications (1)

Publication Number Publication Date
WO2021149528A1 true WO2021149528A1 (ja) 2021-07-29

Family

ID=76992209

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2021/000606 Ceased WO2021149528A1 (ja) 2020-01-22 2021-01-12 イベント予測システム、イベント予測方法およびプログラム

Country Status (3)

Country Link
US (1) US20230058585A1 (https=)
JP (1) JP7440938B2 (https=)
WO (1) WO2021149528A1 (https=)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20220284277A1 (en) * 2021-02-25 2022-09-08 International Business Machines Corporation Network of tensor time series
CN115856208A (zh) * 2022-11-17 2023-03-28 山西中谷科技股份有限公司 一种转动装置转轴运行及状态检测方法、终端和系统
WO2023149236A1 (ja) * 2022-02-04 2023-08-10 国立大学法人大阪大学 学習装置、予測装置、学習方法、予測方法、学習プログラム及び予測プログラム
CN118661139A (zh) * 2022-04-18 2024-09-17 三菱电机株式会社 模拟用程序、模拟装置及控制方法

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20220358182A1 (en) * 2021-05-07 2022-11-10 International Business Machines Corporation Scalable error mitigation
CN116723078A (zh) * 2023-07-14 2023-09-08 中国电信股份有限公司 云服务故障预告警方法及装置
CN117493068B (zh) * 2024-01-03 2024-03-26 安徽思高智能科技有限公司 一种微服务系统根因定位方法、设备及存储介质
CN118656665B (zh) * 2024-08-19 2024-11-05 山东康吉诺技术有限公司 基于深度学习模型的风电机组齿轮箱轴承温度状态检测方法
CN119809404B (zh) * 2024-11-21 2025-10-17 国网上海市电力公司 一种可中断负荷需求响应潜力评价方法
CN120430882B (zh) * 2025-07-08 2025-11-28 西北工业大学 一种社交网络事件预测方法及装置

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2005251185A (ja) * 2004-02-05 2005-09-15 Toenec Corp 電気設備診断システム
WO2018012487A1 (ja) * 2016-07-12 2018-01-18 国立大学法人熊本大学 予測装置、パラメータ集合生産方法及びプログラム
JP2019003389A (ja) * 2017-06-15 2019-01-10 株式会社 日立産業制御ソリューションズ 異常診断装置、異常診断方法及び異常診断プログラム

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2964507B2 (ja) * 1989-12-12 1999-10-18 松下電器産業株式会社 Ｈｍｍ装置
JP6276732B2 (ja) * 2015-07-03 2018-02-07 横河電機株式会社 設備保全管理システムおよび設備保全管理方法
US10504036B2 (en) * 2016-01-06 2019-12-10 International Business Machines Corporation Optimizing performance of event detection by sensor data analytics

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2005251185A (ja) * 2004-02-05 2005-09-15 Toenec Corp 電気設備診断システム
WO2018012487A1 (ja) * 2016-07-12 2018-01-18 国立大学法人熊本大学 予測装置、パラメータ集合生産方法及びプログラム
JP2019003389A (ja) * 2017-06-15 2019-01-10 株式会社 日立産業制御ソリューションズ 異常診断装置、異常診断方法及び異常診断プログラム

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
IRIFUNE, YASUAKI ET AL.: "Detailed labeling to time series data using weak labels", THE 11TH FORUM ON DATA ENGINEERING AND INFORMATION MANAGEMENT (THE 17TH ANNUAL CONFERENCE OF THE DATABASE SOCIETY OF JAPAN, 6 March 2019 (2019-03-06) *
YAMAMURO, SAERU ET AL.: "Summary and Classification of Time Series Data Using Deep Learning", THE 10TH FORUM ON DATA ENGINEERING AND INFORMATION MANAGEMENT (THE 16TH ANNUAL CONFERENCE OF THE DATABASE SOCIETY OF JAPAN, 6 March 2018 (2018-03-06) *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20220284277A1 (en) * 2021-02-25 2022-09-08 International Business Machines Corporation Network of tensor time series
US12561555B2 (en) * 2021-02-25 2026-02-24 International Business Machines Corporation Network of tensor time series
WO2023149236A1 (ja) * 2022-02-04 2023-08-10 国立大学法人大阪大学 学習装置、予測装置、学習方法、予測方法、学習プログラム及び予測プログラム
CN118661139A (zh) * 2022-04-18 2024-09-17 三菱电机株式会社 模拟用程序、模拟装置及控制方法
CN115856208A (zh) * 2022-11-17 2023-03-28 山西中谷科技股份有限公司 一种转动装置转轴运行及状态检测方法、终端和系统

Also Published As

Publication number Publication date
US20230058585A1 (en) 2023-02-23
JPWO2021149528A1 (https=) 2021-07-29
JP7440938B2 (ja) 2024-02-29

Similar Documents

Publication Publication Date Title
JP7440938B2 (ja) イベント予測システム、イベント予測方法およびプログラム
Lindemann et al. A survey on anomaly detection for technical systems using LSTM networks
Esteban et al. Data mining in predictive maintenance systems: A taxonomy and systematic review
Hosseini et al. An ensemble of cluster-based classifiers for semi-supervised classification of non-stationary data streams
Zhong et al. Adaptive memory broad learning system for unsupervised time series anomaly detection
Rodriguez et al. Detecting performance anomalies in scientific workflows using hierarchical temporal memory
Ture et al. Stacking-based ensemble learning for remaining useful life estimation: BA Ture et al.
Ahmadi et al. Modeling recurring concepts in data streams: a graph-based framework
CN114819175A (zh) 人工智能优化平台
Hryniewicz et al. Bayesian analysis of time series using granular computing approach
Alagarsundaram et al. A short-term load forecasting model using restricted Boltzmann machines and bi-directional gated recurrent unit
Kang Product failure prediction with missing data using graph neural networks
Turgunbaev Machine Learning and Its Use in the Automatic Extraction of Metadata from Academic Articles
Kaushik et al. SENSE: software effort estimation using novel stacking ensemble learning: A. Kaushik et al.
Heyden et al. Adaptive Bernstein change detector for high-dimensional data streams
Vishwakarma et al. Taming false positives in out-of-distribution detection with human feedback
Duarte et al. Ensembles of adaptive model rules from high-speed data streams
Sun et al. Time series classification of dynamical systems using deterministic learning
Chen et al. Remaining useful life prediction of milling cutters based on long-term data sequence and parallel fully convolutional feature learning
Wenig et al. Series2graph++: Distributed detection of correlation anomalies in multivariate time series
Kotenko et al. Anomaly detection in iot networks based on intelligent security event correlation
Kaur et al. Remaining useful life improvement for electrical machines using ensemble learning technique
Shahad et al. Challenges in streaming data analysis for building an adaptive model for handling concept drifts
Shivappriya et al. Performance Analysis of Deep Neural Network and Stacked Autoencoder for Image Classification
Talakh et al. Balancing efficiency and accuracy: incremental learning as a key to big data processing

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21744039

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2021573071

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21744039

Country of ref document: EP

Kind code of ref document: A1