Information processing device, analysis method, control system, and analysis program

Abstract

To generate a dynamic model that accurately captures a change in a temporal direction and a spatial direction of an event to be analyzed with a relatively small amount of computation.SOLUTION: An information processing device (1) comprises: a reconstruction unit (103) that generates reconstructed data by reconstructing time-series data (111) from a part of frequency components of the time-series data which is obtained by observing an event to be analyzed in a plant; and an analysis unit (104) that performs spatio-temporal analysis on the reconstructed data to generate a dynamic model (112).SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] This invention relates to an information processing device, etc., that models an event to be analyzed using observational data from a plant. [Background technology]

[0002] Accurately understanding the state of a plant and performing appropriate control according to that state is one of the most important aspects of plant operation. For this reason, the development of technologies for accurately understanding the state of a plant has been progressing for some time. For example, Patent Document 1 below discloses a technology for deriving the position of the burnout point, which is the boundary between the combustion area and the incinerated ash, based on imaging data captured by a stereo camera and the installation position of the stereo camera, in a waste treatment plant equipped with a stoker-type incinerator. [Prior art documents] [Patent Documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2018-155411 [Overview of the Initiative] [Problems that the invention aims to solve]

[0004] It takes time for the results of the control performed to adjust the position of the burnout point to be reflected in the burnout point's position. For this reason, it is desirable to predict the future position of the burnout point, but the technology described in Patent Document 1 can only derive the current position of the burnout point and cannot predict the position of the burnout point in the near future.

[0005] Here, spatiotemporal analysis is a known technique for analyzing time-series observational data. Using spatiotemporal analysis, it is possible to generate a dynamic model that shows the temporal and spatial changes of a target event from time-series data obtained by observing that event.

[0006] However, time-series data observed in plants typically contains a large amount of noise, which degrades the accuracy of dynamic models and increases the computational load and time required to generate them. For example, when the inside of an incinerator is captured as a video, as in Patent Document 1, various elements such as flame fluctuations and smoke movement are captured in addition to the point of complete combustion, and these contribute to a decrease in the accuracy of the dynamic model and an increase in computational load.

[0007] One aspect of the present invention aims to provide an information processing device, etc., that can generate a dynamic model that accurately captures the temporal and spatial changes of an event under analysis with a relatively small amount of computation. [Means for solving the problem]

[0008] To solve the above problems, an information processing device according to one aspect of the present invention comprises a reconstruction unit that generates reconstructed data by reconstructing time-series data from some frequency components of time-series data obtained by observing an event to be analyzed in a plant, and an analysis unit that performs spatiotemporal analysis of the reconstructed data to generate a dynamic model showing the temporal and spatial changes of the event.

[0009] Furthermore, an analysis method according to one aspect of the present invention is an analysis method performed by one or more information processing devices in order to solve the above problems, comprising: a reconstruction step of generating reconstructed data by reconstructing time-series data from some frequency components of time-series data obtained by observing an event to be analyzed in a plant; and an analysis step of generating a dynamic model showing the temporal and spatial changes of the event by performing spatiotemporal analysis on the reconstructed data. [Effects of the Invention]

[0010] According to one aspect of the present invention, it becomes possible to generate a dynamic model that accurately captures the temporal and spatial changes of the event being analyzed with a relatively small amount of computation. [Brief explanation of the drawing]

[0011] [Figure 1] It is a diagram showing an example of the main part configuration of an information processing apparatus according to an embodiment of the present invention. [Figure 2] It is a diagram showing a configuration example of a control system including the above information processing apparatus. [Figure 3] It is a diagram showing an outline of an analysis method according to an embodiment of the present invention. [Figure 4] It is a diagram showing an example in which a dynamic model is generated from time-series data extracted from a moving image taken inside an incinerator, and a predicted image is generated by the generated dynamic model. [Figure 5] It is a flowchart showing an example of the processing executed by the above information processing apparatus. [Figure 6] It is a flowchart showing another example of the processing executed by the above information processing apparatus.

Mode for Carrying Out the Invention

[0012] 〔System Configuration〕 The configuration of a control system 7 according to an embodiment of the present invention will be described based on FIG. 2. FIG. 2 is a diagram showing a configuration example of the control system 7. The control system 7 is a system for controlling various devices provided in an incinerator in a waste incineration plant. This incinerator is configured to incinerate waste supplied from a hopper shown in (a) in FIG. 2 while conveying it by a fire grate shown in (b) in a furnace shown in (c). Also, an operator can control various devices such as a fire grate from a control room shown in (d). The waste may be any combustible waste, for example, general waste such as household waste or industrial waste.

[0013] As shown in the figure, the control system 7 includes an information processing device 1, a control device 2, and a photographing device 3. In the example of FIG. 2, among these components, the information processing device 1 and the control device 2 are provided in the control room, and the photographing device 3 is provided at a position where the inside of the furnace can be photographed (more specifically, at a position on the downstream side in the waste conveyance direction). Note that the information processing device 1 and the control device 2 may be provided outside the control room.

[0014] Details will be described later. The information processing device 1 generates reconstructed data by reconstructing the time-series data from some frequency components of the time-series data obtained by observing an event to be analyzed in the plant. Then, the information processing device 1 performs spatio-temporal analysis on the generated reconstructed data to generate a dynamic model showing changes in the time direction and the space direction of the above event, and makes inferences about the above event using this dynamic model.

[0015] In the present embodiment, an example will be described in which the event to be analyzed is the combustion of waste in the incinerator in the waste incineration plant. Also, in the present embodiment, an example will be described in which the time-series data is time-series frame images extracted from a moving image showing the state of combustion of waste in the incinerator, photographed by the photographing device 3. Therefore, the above dynamic model is a model showing changes in the time direction and the space direction of the combustion state of waste in the incinerator.

[0016] The control device 2 controls the plant equipment based on the result of the inference by the information processing device 1. The equipment to be controlled may be any equipment that affects the combustion state of waste in the incinerator. For example, the control device 2 may perform control to adjust the grate speed (which can also be referred to as the waste conveyance speed) by controlling the equipment that operates the grate, or control to adjust the supply amount of combustion air by controlling the combustion air supply device.

[0017] As described above, the control system 7 includes an information processing device 1 that generates reconstructed data by reconstructing the time series data from some of the frequency components of the time series data obtained by observing the event under analysis, generates a dynamic model by performing spatiotemporal analysis on the generated reconstructed data, and performs inferences regarding the above event using the generated dynamic model, and a control device 2 that controls the equipment of the plant based on the results of the inference by the information processing device 1.

[0018] As will be explained in detail below, with the above configuration, compared to directly performing spatiotemporal analysis on the time-series data being analyzed, it is possible to obtain highly accurate inference results using a dynamic model that accurately captures the temporal and spatial changes of some frequency components with less computation. Furthermore, it is possible to automatically perform appropriate control based on these results. For example, if the control device 2 determines through inference by the information processing device 1 that the burnout point is shifting downstream beyond a reasonable range, it may perform control to reduce the grate velocity. This makes it possible to keep the burnout point within a reasonable range.

[0019] [Configuration of the information processing device] The configuration of the information processing device 1 will be explained based on Figure 1. Figure 1 is a block diagram showing an example of the main components of the information processing device 1. As shown in the figure, the information processing device 1 includes a control unit 10 that controls all parts of the information processing device 1, and a storage unit 11 that stores various data used by the information processing device 1. The information processing device 1 also includes a communication unit 12 for the information processing device 1 to communicate with other devices, an input unit 13 that receives input of various data to the information processing device 1, and an output unit 14 for the information processing device 1 to output various data.

[0020] The control unit 10 also includes a data acquisition unit 101, a conversion unit 102, a reconstruction unit 103, an analysis unit 104, and an inference unit 105. The storage unit 11 stores time-series data 111 and a dynamic model 112.

[0021] The data acquisition unit 101 acquires time-series data obtained by observing the events under analysis in the plant. The acquired time-series data is stored in the storage unit 11 as time-series data 111. Specifically, the data acquisition unit 101 acquires a video image showing the combustion of waste inside the incinerator, captured by the imaging device 3 shown in Figure 2, and acquires time-series data 111 by extracting frame images from this video image at a predetermined sampling period. The data acquisition unit 101 may also resize the frame images before using them as time-series data 111.

[0022] The conversion unit 102 converts the time series data 111 into frequency domain data (which can also be called frequency components). For example, the conversion unit 102 may convert the time series data 111 into frequency domain data by performing a Discrete Fourier Transform (DFT) or a Fast Fourier Transform (FFT).

[0023] The reconstruction unit 103 generates reconstructed data by reconstructing the time-series data 111 from some of its frequency components. Specifically, when the reconstruction unit 103 inversely transforms the frequency-domain data generated by the transformation unit 102 back into time-domain data, it uses only a portion of the frequency-domain data, rather than all of it, to generate the reconstructed data. This process can be described as removing unnecessary frequency components or extracting necessary frequency components.

[0024] The analysis unit 104 performs spatiotemporal analysis on the reconstruction data generated by the reconstruction unit 103 to generate a dynamic model that shows the temporal and spatial changes of the event under analysis. The generated dynamic model is stored in the storage unit 11 as the dynamic model 112. As described above, in this embodiment, an example is described in which the dynamic model 112 is a model that shows the temporal and spatial changes of the combustion state of waste in the incinerator.

[0025] The inference unit 105 uses the dynamic model 112 to perform inferences regarding the events under analysis in the plant. In this embodiment, an example is described in which the inference unit 105 uses the dynamic model 112 to generate a predictive image showing the future combustion state of waste in the incinerator.

[0026] As described above, the information processing device 1 includes a reconstruction unit 103 that generates reconstructed data by reconstructing the time series data 111 from some of the frequency components of the time series data 111 obtained by observing the events to be analyzed in the plant, and an analysis unit 104 that performs spatiotemporal analysis on the reconstructed data to generate a dynamic model 112 that shows the changes in the time and spatial directions of the above events.

[0027] According to the above configuration, instead of directly performing spatiotemporal analysis on the time series data being analyzed, the reconstructed data, which is created by reconstructing the time series data from some of its frequency components, is then subjected to spatiotemporal analysis. Therefore, compared to directly performing spatiotemporal analysis on the time series data being analyzed, it becomes possible to generate a dynamic model that accurately captures the temporal and spatial changes of the event being analyzed with less computation.

[0028] Furthermore, the information processing device 1 includes an inference unit 105 that performs inferences about the event under analysis using a dynamic model 112. This makes it possible to obtain highly accurate inference results from the dynamic model 112, which accurately captures changes in the time and spatial directions of some frequency components. The content of the inference is not particularly limited; for example, the inference unit 105 may generate a predictive image showing the future combustion state of the waste. The inference unit 105 may also analyze the generated predictive image and classify the future combustion state of the waste.

[0029] [Overview of the analysis method] The overview of the analysis method of this embodiment will be explained with reference to Figure 3. Figure 3 is a diagram illustrating the overview of the analysis method of this embodiment. In the analysis method of this embodiment, spatiotemporal analysis is performed using a technique called DMD (Dynamic Mode Decomposition). Figure 3 shows time-series data 111, which are frame images extracted from video footage taken inside an incinerator, a group of graphs 31 showing each frequency component obtained by converting the time-series data 111 into the frequency domain, and a group of images 32 showing the dynamic modes obtained by DMD.

[0030] Graphs 31a to 31c, included in graph group 31, each show different frequency components. Of the frequency components shown in these three graphs, the frequency component shown in graph 31a is the lowest frequency, and the frequency component shown in graph 31c is the highest frequency. In video footage taken inside an incinerator, changes such as flame flicker appear as high-frequency components, while changes in combustion position appear as low-frequency components. In monitoring waste incineration plants, low-frequency components that show changes over relatively long periods, such as 10 minutes or more, are important.

[0031] Each image in image group 32 represents the dynamic mode determined by DMD as an image of the inside of the incinerator. Figure 3 shows a graph illustrating the waveform of the dynamic mode corresponding to each image. In this way, DMD analyzes the event being analyzed by decomposing the time-series data 111 into multiple dynamic modes. This analysis method is explained below.

[0032] In DMD, the time series change of data is represented by the linear operator A, as shown in equation (1) below. That is, the state x at a certain time point k+1. k+1 The state x at the time k immediately before that point. k It is expressed as the product of and the linear operator A. For example, if the time series data 111 is a frame image extracted from a video as shown in Figure 3, then the following x k The vector below represents the pixel value (e.g., RGB value) of each frame image, K represents the number of frame images, and P represents the resolution of each frame image.

[0033]

number

[0034]

number

[0035] In DMD, as shown in equation (2) below, X1 = [x1, ..., x k-1 ] and X2=[x2,…,x k We seek the operator A that minimizes the error between snapshot matrices whose time series are shifted by 1. A snapshot is a vector representing the state at a given time. F represents the Frobenius norm.

[0036]

number

[0037] The above A can be calculated using the least squares method, as shown in equation (3) below. In equation (3), the symbol superscript to the right of X1 represents the Moore-Penrose pseudoinverse.

[0038]

number

[0039] Although details are omitted, in DMD, the eigenvalues ​​and eigenvectors of operator A are obtained by eigenvalue decomposition of the operator A described above. The eigenvectors of A are the dynamic modes. As shown in the image group 32 in Figure 3, numerous dynamic modes are calculated from the time series data 111.

[0040] As described above, the graph group 31 shown in Figure 3 represents the frequency components of the time-series data 111. Among the graphs included in graph group 31, the frequency components of graphs 31a to 31c correspond to the dynamic modes of images 32a to 32c included in image group 32, respectively. Thus, a correspondence exists between the frequency components obtained by performing a Fast Fourier Transform or Discrete Fourier Transform on the time-series data 111 and the dynamic modes calculated by DMD.

[0041] The computational complexity of DMD increases as the number of dynamic modes N to be calculated increases. Therefore, in the analysis method of this embodiment, reconstructed data is generated by reconstructing the time series data 111 from some of the frequency components of the time series data 111. For example, if we want to find the dynamic mode shown in image 32b among the dynamic modes shown in image group 32, we generate reconstructed data using the frequency components corresponding to graph 31b, which corresponds to image 32b, from the graphs included in graph group 31. This makes it possible to calculate only the dynamic mode shown in image 32b among the dynamic modes shown in image group 32 at high speed with a small amount of computation.

[0042] Specifically, in the analysis method of this embodiment, first, the conversion unit 102 converts the time series data 111 into frequency domain data. Here, x in equation (1) k This applies when the object being measured at each time step includes an element of P.

[0043]

number

[0044] This is the result. k From sampling frequency f s One-dimensional time series data sampled

[0045]

number

[0046] with the sampling frequency being s = [0, 1, 2, …, (K - 1)] / f s and converting it to data in the frequency domain with J = K

[0047] [Number]

[0048] is obtained. And each frequency component y j is the weighted sum of all elements included in x, that is, represented by the following formula (4).

[0049] [Number]

[0050] The conversion unit 102 may, for example, convert the time-series data 111 into data in the frequency domain according to the above formula (4). This conversion is the same as a general discrete Fourier transform.

[0051] Next, the reconstruction unit 103 inversely converts the data in the frequency domain obtained by the above conversion according to, for example, the following formula (5) to return it to an image. As a result, reconstruction data corresponding to some frequency components of the time-series data 111 is obtained. The reconstruction data has the same number of pixels and size as the original time-series data 111, but information regarding some frequency components is missing. Therefore, it can be said that the reconstruction data is approximate data with the rank of the original time-series data 111 reduced.

[0052] [Number]

[0053] The above formula (5) is obtained by adding ζ to the formula of a general inverse Fourier transform. ζ is, as shown below, a specific y jThis is an indicator function that determines whether or not to include it in the reconstructed data. Furthermore, τ represents domain knowledge related to the task. Specifically, τ represents the task-related frequencies, where τ = {τ1, τ2, ..., τ D The task in this embodiment is to generate a dynamic model that shows the temporal and spatial changes of the event under analysis in the plant over a relatively long period (e.g., 10 minutes or more).

[0054]

number

[0055] Task-related frequency information τ can be set to one value or multiple values. For example, in the example in Figure 3, setting τ to represent the frequency components shown in graph 31b will calculate the dynamic mode corresponding to image 32b. Also, setting τ (a total of three τ) corresponding to each frequency component shown in graphs 31a to 31c will calculate the dynamic modes corresponding to images 32a to 32c.

[0056] The value of task-related frequency information τ can be predetermined according to the event being analyzed. For example, the value of τ may be set based on the operating cycle of equipment operating in the plant. In the incinerator of a waste incineration plant, the grate moves periodically, so the operating cycle of this grate can be used as a reference, and frequency components with frequencies higher than that operating cycle may not be used, while frequency components with frequencies lower than that operating cycle may be used. The inventors of this invention have set τ = 300 -1 Experiments using Hz have successfully enabled accurate estimation of the conditions inside the incinerator.

[0057] Thus, the reconstruction unit 103 may generate reconstruction data from frequency components selected based on the operating cycle of the equipment operating in the plant. This makes it possible to generate reconstruction data from appropriate frequency components that take into account the operation of the equipment in the plant.

[0058] The process of determining task-related frequency information τ based on the operating cycle of the equipment operating in the plant may be performed by the user of the information processing device 1, or by the information processing device 1 itself. In the latter case, the information processing device 1 accepts input of the operating cycle of the equipment and automatically sets the value of τ according to the input operating cycle. Furthermore, if the operating cycle of the equipment operating in the plant is changed, the reconstruction unit 103 may update τ to a value corresponding to the changed operating cycle.

[0059] The reconstructed data generated by the above process is represented as follows.

[0060]

number

[0061] Then, the analysis unit 104 calculates the eigenvalues ​​and eigenvectors of operator A using DMD with the above-mentioned reconstructed data. In this case, the above formula (3) can be rewritten as follows.

[0062]

number

[0063] The dynamic model 112 generated by the analysis unit 104 is calculated using the eigenvalue Λ calculated by DMD. t It can be expressed by the following equation (6) using the eigenvector (dynamic mode) Φ. Here, b is the amplitude of the corresponding dynamic mode. Furthermore, since the eigenvalues, eigenvectors, and amplitudes shown in equation (6) were calculated using task-related frequency information τ, the letter τ is appended to each of these corresponding letters.

[0064]

number

[0065] After the dynamic model 112 shown by the above formula (6) is generated as described above, the inference unit 105 uses this dynamic model 112 to determine the state x at any given time t. t It can be predicted.

[0066] Thus, learning in DMD is completed by calculating the eigenvalues ​​and eigenvectors of operator A. Therefore, compared to machine learning methods that require a large amount of training data, such as neural networks, it has the advantage of not requiring the effort of collecting training data or labeling the data for training. It also has the advantage of being applicable to modeling phenomena where it is difficult to collect a large amount of training data.

[0067] [Specific examples of processing] An example of applying the process described above to time-series data 111 extracted from video footage taken inside an incinerator will be explained with reference to Figure 4. Figure 4 shows an example in which a dynamic model 112 is generated from time-series data 111 extracted from video footage taken inside an incinerator, and a predicted image is generated using the generated dynamic model 112.

[0068] In the example in Figure 4, x1 to x k The input data is time-series data 111 consisting of k frame images. The conversion unit 102 converts x1 to x included in this time-series data 111. k Each of these is transformed into frequency domain data by a Fourier transform (e.g., FFT). Then, the reconstruction unit 103 reconstructs the time series data 111 from some of the frequency components contained in the frequency domain data (frequency components extracted based on task-related frequency information τ) by an inverse Fourier transform (e.g., inverse FFT). The Fourier transform and inverse Fourier transform may be performed using the above-mentioned equations (4) and (5), respectively.

[0069] Next, the analysis unit 104 performs spatiotemporal analysis of the reconstructed data using DMD to calculate the dynamic mode Φ and eigenvalues ​​Λ. The τ used in these calculations is given by τ = {τ1, τ2, ..., τ DIf we assume}, the calculated dynamic modes are Φ={Φ1,Φ2,…,Φ D This results in D of}. Figure 4 shows image 41, which visualizes these dynamic modes.

[0070] As described above, the eigenvalue Λ and dynamic mode Φ are calculated, generating the dynamic model 112 shown in equation (6). In the example in Figure 4, the inference unit 105 uses this dynamic model 112 to generate predictive images 42 showing the state inside the incinerator at time t and t+1, respectively.

[0071] Predicted image 42 is the time series data 111x1~x used to calculate the eigenvalues ​​and dynamic modes. k This indicates the state of the incinerator at a point in time further ahead, i.e., in the future. As long as the trend of changes in the incinerator's state is maintained, the prediction using the calculated eigenvalues ​​and dynamic modes remains valid. When the trend of changes in state changes, the eigenvalues ​​and dynamic modes can be recalculated using the time series data 111 acquired after the change.

[0072] Furthermore, by using the dynamic model 112, it is possible to generate a predicted image for the time series data 111 at the time it was observed. By comparing this predicted image, i.e., the sequence generated using the dynamic model 112, with the time series data 111 at the corresponding time point, i.e., the original sequence, the prediction accuracy can be evaluated. Experiments by the inventors of this application have confirmed that it is possible to generate a predicted image with sufficient accuracy for practical use.

[0073] Furthermore, while prediction images are discrete data, generating time-series prediction images allows us to predict the rate of state change from those prediction images. Additionally, DMD calculates information representing the periodicity of changes in each dynamic mode (temporal dynamics). This periodicity can be said to indicate the rate of change.

[0074] [Processing flow] The flow of processing (analysis method) performed by the information processing device 1 will be explained based on Figure 5. Figure 5 is a flowchart showing an example of processing performed by the information processing device 1.

[0075] In S11, the data acquisition unit 101 acquires a video image taken by the camera 3 showing the burning of waste inside the incinerator. Then, in S12, the data acquisition unit 101 acquires time-series data by extracting frame images from the video image acquired in S11 and stores it in the storage unit 11 as time-series data 111.

[0076] In S13, the conversion unit 102 converts the time-series data 111 stored in S12 into frequency-domain data. For this conversion, for example, the above formula (4) may be used.

[0077] In S14 (reconstruction step), the reconstruction unit 103 reconstructs time-series data 111 from some of the frequency components that make up the frequency domain data obtained by the conversion in S13, and generates reconstructed data. For this reconstruction, for example, the above formula (5) may be used. In this case, task-related frequency information τ indicating the frequency components to be used is used.

[0078] In step S15 (analysis step), the analysis unit 104 generates a dynamic model 112 by performing spatiotemporal analysis on the reconstructed data generated in S14. For example, the analysis unit 104 may generate the dynamic model 112 shown in equation (6) by calculating the eigenvalues ​​and eigenvectors of operator A in equation (3)' described above.

[0079] In S16, the inference unit 105 uses the dynamic model 112 generated in S15 to perform inferences about the events under analysis in the plant. For example, the inference unit 105 uses the dynamic model 112 shown by equation (6) to determine the state x at any given time t. tThis can also be predicted. This completes the process shown in Figure 5. After S16, the control device 2 may control the equipment in the plant based on the inference result of S16. It is necessary to predetermine what kind of control to perform on which equipment when a certain inference result is obtained.

[0080] As described above, the analysis method according to this embodiment includes a reconstruction step (S14) which generates reconstructed data by reconstructing the time series data 111 from some frequency components of the time series data 111 obtained by observing the event to be analyzed in the plant, and an analysis step (S15) which generates a dynamic model 112 that shows the temporal and spatial changes of the event to be analyzed by performing a spatiotemporal analysis on the reconstructed data. This makes it possible to generate a dynamic model 112 that accurately captures the temporal and spatial changes of the event to be analyzed with a relatively small amount of computation.

[0081] [Processing flow (real-time processing)] The reconstruction unit 103 may generate reconstructed data from the time-series data 111 each time additional time-series data 111 is added. The analysis unit 104 may then update the dynamic model 112 using the newly generated reconstructed data, and the inference unit 105 may perform inferences about the event under analysis using the updated dynamic model 112. This makes it possible to output inference results in real time or at a very fast pace.

[0082] In this case, the analysis unit 104 may generate and update the dynamic model 112 using a method called, for example, streaming DMD. Streaming DMD is a method that updates the dynamic model 112 whenever a new snapshot becomes available. For specific details of the process, see, for example, "Dynamic Mode Decomposition for Large and Streaming Datasets" by Maziar S. Hemati et al., Physics of Fluids, vol. 26, no. 11, p. 111701, 2014.

[0083] The flow of processing (analysis method) executed by the information processing device 1 when performing the above-described processing will be explained based on Figure 6. Figure 6 is a flowchart of an example of other processing performed by the information processing device 1. Processes similar to those in Figure 6 are given the same number. These processes will not be explained again.

[0084] In S15A, the analysis unit 104 generates a dynamic model 112 by performing spatiotemporal analysis on the reconstructed data generated in S14. For example, the analysis unit 104 generates the dynamic model 112 using the streaming DMD described above. If a previously generated dynamic model 112 already exists, the analysis unit 104 updates the previously generated dynamic model 112 using the reconstructed data generated in the most recent S14.

[0085] In S17A, the data acquisition unit 101 determines whether or not to terminate the analysis process. If it is determined to terminate (YES in S17A), the analysis process ends. On the other hand, if it is determined not to terminate (NO in S17A), the process returns to S11. This allows the latest inference results to be output in real time or at a very fast pace while the video is being captured. Note that the termination condition in S17A is not particularly limited; for example, the termination condition may be the completion of video capture.

[0086] [Other application examples] In the embodiments described above, the focus was on an example of generating a dynamic model 112 that shows the combustion state of waste inside the incinerator of a waste incineration plant. However, it is also possible to generate dynamic models that represent other events in a waste incineration plant. Furthermore, it is also possible to generate dynamic models that represent events in plants other than waste incineration plants.

[0087] For example, in a sludge treatment plant, a coagulant is added to the sludge and stirred to separate solid components from the sludge. In this process, by analyzing images of the stirred sludge using the analysis method described in the above embodiment, a dynamic model showing the temporal and spatial state changes of the sludge can be generated. Using such a dynamic model, it becomes possible to predict the future state of the sludge and appropriately control the timing and amount of coagulant addition based on the prediction results. In this case, the task-related frequency information τ can be set to a value corresponding to the operating cycle of the stirring device that stirs the sludge, for example.

[0088] [Variation] The entity executing each process described in the above-described embodiment is arbitrary and is not limited to the examples above. For example, processes S13 to S15 among the processes shown in Figure 5 may be executed by a device separate from the information processing device 1 (for example, a cloud server). In this case, the information processing device 1 only needs to transmit the time-series data to the above device and obtain the dynamic model generated by the device. Thus, the analysis method according to this embodiment can be executed by one information processing device 1 or by multiple information processing devices.

[0089] Furthermore, although the information processing device 1 and the control device 2 are separate devices in the control system 7 shown in Figure 2, the same functions as the control system 7 can be achieved by integrating them into a single device.

[0090] Furthermore, while DMD and streaming DMD were described as examples of spatiotemporal analysis methods in the embodiments described above, the spatiotemporal analysis methods applicable to the present invention are not limited to these examples.

[0091] Furthermore, although the above-described embodiment explained an example in which frame images extracted from moving images are used as time-series data, the time-series data used in the present invention is not limited to this example and can be obtained by observing the events to be analyzed in the plant. For example, time-series measurements from various sensors installed at various locations in the plant may be used as time-series data. In this case, a dynamic model is generated that shows the changes in the time and spatial directions of the events to be analyzed, corresponding to the spatial position of each sensor and its measured value.

[0092] [Examples of implementation using software] The function of the information processing device 1 is an analysis program for causing the computer to function as the information processing device 1, and this can be realized by an analysis program for causing the computer to function as each control block of the information processing device 1 (particularly each part included in the control unit 10).

[0093] In this case, the information processing device 1 includes a computer having at least one control device (e.g., a processor) and at least one storage device (e.g., memory) as hardware for executing the analysis program. By executing the analysis program using this control device and storage device, each of the functions described in the above embodiment is realized.

[0094] The analysis program described above may be recorded on one or more computer-readable recording media, rather than temporarily. These recording media may or may not be provided by the information processing device 1. In the latter case, the analysis program may be supplied to the device via any wired or wireless transmission medium.

[0095] Furthermore, some or all of the functions of each of the above control blocks can also be realized by logic circuits. For example, an integrated circuit in which logic circuits functioning as each of the above control blocks are formed is also included in the scope of the present invention. In addition, it is also possible to realize the functions of each of the above control blocks by, for example, a quantum computer.

[0096] The present invention is not limited to the embodiments described above, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention. [Explanation of Symbols]

[0097] 1. Information Processing Device 103 Reconstruction part 104 Analysis Department 105 Reasoning part 111 Time series data 112 Dynamic Models 2 Control device 7 Control System

Claims

1. A reconstruction unit generates reconstructed data by reconstructing time-series data from some of the frequency components of time-series data obtained by observing the events under analysis in a plant, The system includes an analysis unit that performs spatiotemporal analysis on the reconstructed data to generate a dynamic model showing the temporal and spatial changes of the event, The reconstruction unit generates the reconstructed data by inversely transforming some of the frequency components. The analysis unit calculates eigenvalues ​​and eigenvectors of a linear operator representing the time-series change of the reconstructed data using the reconstructed data, and generates the dynamic model represented by the calculated eigenvalues ​​and eigenvectors, which is an information processing device.

2. The information processing device according to claim 1, further comprising an inference unit that performs inferences regarding the events using the dynamic model.

3. The reconstruction unit generates the reconstructed data from the time-series data each time the time-series data is added, The analysis unit updates the dynamic model using the newly generated reconstruction data. The information processing apparatus according to claim 2, wherein the inference unit performs inference regarding the event using the updated dynamic model.

4. The information processing apparatus according to any one of claims 1 to 3, wherein the reconstruction unit generates the reconstruction data from frequency components selected based on the operating cycle of the equipment operating in the plant.

5. An analysis method performed by one or more information processing devices, A reconstruction step that generates reconstructed data by reconstructing the time series data from some of the frequency components of the time series data obtained by observing the event under analysis in the plant, The analysis step includes performing a spatiotemporal analysis on the reconstructed data to generate a dynamic model that shows the temporal and spatial changes of the event, In the reconstruction step, the reconstructed data is generated by inversely transforming some of the frequency components. The analysis method, in the analysis step, calculates eigenvalues ​​and eigenvectors of a linear operator representing the time-series change of the reconstructed data using the reconstructed data, and generates the dynamic model represented by the calculated eigenvalues ​​and eigenvectors.

6. The information processing apparatus according to claim 2, A plant control system including a control device that controls the equipment of the plant based on the results of the aforementioned inference.

7. An analysis program for causing a computer to function as an information processing device according to claim 1, wherein the reconstruction unit and the analysis unit are the computer.

Images