JP7308476B2 - Information processing device, information processing method, and information processing program - Google Patents

Description

TECHNICAL FIELD The present invention relates to an information processing device and the like for predicting the future state of a plant based on data measured in the plant.

2. Description of the Related Art Conventionally, in a plant such as a waste incineration plant, research has been conducted to predict the future state of the plant from various process data such as sensor detection values. In conventional prediction, prediction using a trained model such as a neural network generated by machine learning is also known. However, sufficient prediction accuracy cannot be obtained simply by machine learning the patterns of fluctuations in past process data values. was left.

Also, as a technique using a trained model, for example, the technique described in the following document is also known. The document below describes an encoder that infers latent variables from input data, a decoder that generates restored data from latent variables, and a determination means for determining whether input data is normal based on the difference between input data and restored data. An anomaly detection system is described.

WO2017/094267

With the technique of the above document, it is possible to determine whether or not the input data is normal, but it is not possible to predict what value the input data will take in the future. Therefore, even if the technique of the above document is used, the future state of the plant cannot be predicted from the data regarding the processes in the plant.

One aspect of the present invention has been made in view of the above problems, and an object thereof is to provide an information processing apparatus or the like capable of predicting the future state of a plant from data relating to processes in the plant. .

In order to solve the above problems, an information processing apparatus according to an aspect of the present invention includes an encoder that encodes time-series data relating to a predetermined process in a plant and outputs a latent variable, and a model that models fluctuations in the latent variable. a prediction unit that predicts the future value of the latent variable from the value of the latent variable output by the encoder using the dynamic model obtained by the dynamic model; and a decoder that outputs a predicted value of the time-series data.

Further, an information processing method according to an aspect of the present invention is an information processing method using an information processing device to solve the above problems, wherein time-series data relating to a predetermined process in a plant is encoded to generate a latent variable. a step of outputting, a prediction step of predicting a future value of the latent variable from the value of the latent variable output by encoding using a dynamic model that models fluctuations in the latent variable, and the prediction step and decoding the value of the latent variable predicted by and outputting the predicted value of the time series data.

According to one aspect of the present invention, future plant conditions can be predicted from data relating to processes in the plant.

1 is a block diagram showing an example of a main configuration of an information processing apparatus according to Embodiment 1 of the present invention; FIG. It is a figure which shows the outline|summary of the prediction by the said information processing apparatus. It is a figure which shows an example of the network structure of VAE. It is a figure which shows the structural example of a state space model. It is a flow chart which shows an example of processing which the above-mentioned information processor performs. It is a figure which shows the outline|summary of the prediction by the information processing apparatus which concerns on Embodiment 2 of this invention. It is a block diagram which shows an example of the plant management system containing the said information processing apparatus.

[Embodiment 1]
〔Device configuration〕
The configuration of the information processing apparatus 1 of this embodiment will be described with reference to FIG. FIG. 1 is a block diagram showing an example of the main configuration of an information processing apparatus 1. As shown in FIG. The information processing device 1 is a device that predicts the future state of the plant from data on processes in the plant. In the following, an example of predicting future process data values and future in-furnace images from process data in the garbage incineration process of a refuse incineration plant and in-furnace images that are images of the inside of the incinerator is taken. explain. However, the plant is not limited to a waste incineration plant, as long as it uses at least one machine or device to carry out a prescribed treatment (apply a prescribed treatment to a prescribed object, manufacture a prescribed article, etc.). do not have. Also, the process is not limited to the garbage incineration process.

The process data may be any data that reflects the state of the garbage incineration process. For example, the values of instruments in the plant and the output values of various sensors installed in the plant may be used as the process data. Further, the in-furnace image may be an image showing the state of the garbage incineration process in chronological order. Alternatively, the in-furnace images may be frame images extracted in chronological order from moving images of the inside of the incinerator.

As shown in the figure, the information processing apparatus 1 includes a control unit 10 for centrally controlling each unit of the information processing apparatus 1, a storage unit 20 for storing various data used by the information processing apparatus 1, and an input unit for receiving input operations. and an output unit 40 for outputting data. The control unit 10 also includes an image preprocessing unit 101 , a PD preprocessing unit 102 , an encoder 103 , a prediction unit 104 and a decoder 105 .

The image preprocessing unit 101 performs predetermined preprocessing on the in-furnace image. It is preferable that the content of the preprocessing be such that the characteristic portion of the in-furnace image is not lost and the characteristic portion can be made clearer. For example, the image preprocessing unit 101 divides the image area of the in-furnace image into an area in which the fire grate in the in-furnace image, garbage burning on the grate, and flames rising from the garbage appear, and an incinerator as a background of these. It is also possible to perform preprocessing to extract the former area by dividing the area into areas such as wall surfaces and the like.

The PD preprocessing unit 102 performs predetermined preprocessing on process data. It is preferable that the content of the preprocessing be such that the characteristic portion of the process data can be made clearer without losing the characteristic portion. In this embodiment, as an example, an example in which the PD preprocessing unit 102 executes preprocessing for removing noise from process data will be described. Noise removal can be performed by, for example, a low-pass filter or the like.

Other examples of preprocessing by the PD preprocessing unit 102 include process data normalization processing, processing for interpolating missing values in process data, processing for removing outliers in process data, and selection of attributes of process data. processing, and processing for dimensionally compressing process data. Note that the attribute selection process includes, for example, a process of grasping feature points (shape, color, etc. in the case of an image) of process data by principal component analysis and emphasizing the feature points. In addition to this, for example, from frequency data obtained by Fourier transforming process data, a process of extracting characteristic points of the process data may be performed as preprocessing. The image preprocessing unit 101 may also perform such preprocessing on the in-furnace image.

Encoder 103 encodes time-series data for a given process in the plant and outputs latent variables. Also, the prediction unit 104 predicts the future value of the latent variable from the value of the latent variable output by the encoder 103 using a dynamic model that models the variation of the latent variable. Then, the decoder 105 decodes the value of the latent variable predicted by the prediction unit 104 and outputs the predicted value of the time-series data.

According to the above configuration, by encoding the time-series data, it is possible to output a latent variable whose dimensionality is reduced from that of the time-series data while retaining the characteristics of the time-series data. Further, according to the above configuration, it is possible to predict the future value of the latent variable and decode the predicted value of the latent variable by using a dynamic model obtained by modeling variations in the latent variable by machine learning. can be used to output forecast values for time-series data. Therefore, according to the above configuration, it is possible to predict the future state of the refuse incineration plant from the data relating to the refuse incineration process in the refuse incineration plant.

More specifically, in this embodiment, the encoder 103 encodes process data and in-furnace images in the garbage incineration process of the garbage incineration plant and outputs latent variables. Also, the prediction unit 104 uses a Linear Gaussian State Space Model (LGSSM) as the dynamic model, and predicts the future value of the latent variable from the value of the latent variable output by the encoder 103. do. Then, the decoder 105 decodes the value of the latent variable predicted by the prediction unit 104 and outputs the predicted value of the process data and the predicted value of the in-core image (hereinafter also referred to as the predicted image).

As described above, in this embodiment, the time-series data used for prediction includes an image of the progress of the process and the process data acquired during the progress of the process. The encoder 103 outputs latent variables reflecting the feature points of the in-furnace image and the feature points of the process data, and the decoder 105 decodes the values of the latent variables predicted by the prediction unit 104, and a predicted image.

The above configuration is not essential, but according to this configuration, the prediction is performed taking into account the feature points of the in-furnace image, so the prediction accuracy of the process data is improved compared to the case where the prediction is performed only from the process data. It is preferable because it can Moreover, according to the above configuration, it is possible to output a predictive image showing the future state of the process, so there is also the advantage that the user can visually grasp the future state of the process.

[Outline of forecast]
An overview of prediction by the information processing device 1 will be described with reference to FIG. FIG. 2 is a diagram showing an overview of prediction by the information processing device 1. As shown in FIG. The information processing device 1 performs prediction using the theory of KVAE (Kalman Variational Autoencoder). This prediction includes an observation step and a prediction step, as shown in FIG.

In the observation step, the encoder 103 receives input of observation data (specifically, noise-removed process data and preprocessed in-furnace image) and outputs latent variables. Then, in the prediction step, the future latent variables predicted by the prediction unit 104 are input to the decoder 105, and the decoder 105 generates predicted values of future process data (noise removed) and preprocessed predicted images. Output. Details of encoding and decoding will be described later with reference to FIG.

Using a state space model, the prediction unit 104 predicts the future state of the waste incineration plant (hidden state ) is predicted. This state space model is a dynamic model that models the variation of the latent variable that is the series of observed values. Then, in the prediction step, the prediction unit 104 converts the state variables into latent variables (future latent variables) using the state space model and outputs the latent variables to the decoder 105 .

Specifically, in the example of FIG. 2, the process data y _t-1 and the in-furnace image x t- ₁ at time (t-1) are input to the encoder 103, and the latent variable a _t-1 is output. , the process data _yt and the in-furnace image _xt at time t are input and the latent variable _at is output.

Since the value _of the state variable z at a certain time depends on the value of the state variable z one time before, as shown in FIG. Input data u _t-1 can be input and the state variable z at time t can be output. Similarly, the latent variable at at time _t and the control input data u _t can be input to the state space model to output the state variable zt ₊₁ at time t+1.

Then, the prediction unit 104 converts the state variable z _t+1 into the latent variable a _t+1 and outputs it to the decoder 105 . Finally, the decoder 105 decodes the latent variable a _{t+1 to} obtain the predicted value of the process data at time (t+1).

and the predicted image at time (t+1) (predicted value of preprocessed in-core image)

is outputting

As shown in FIG. 2, the prediction unit 104 uses control input data indicating the content of the control input to the waste incineration plant, and from the value of the latent variable output by the encoder 103, the control input indicated by the control input data is determined. It can predict the value of the latent variable after it is done. This kind of prediction is based on a dynamic model (Fig. This is possible by using the state space model in the example of 2).

Although the above configuration is not essential, this configuration is preferable because it is possible to predict the value of the latent variable after the control input indicated by the control input data is performed. Since the predicted value of the latent variable is decoded by the decoder 105, according to the above configuration, the state of the refuse incineration plant after the control input is performed can be predicted. This prediction result can be used, for example, as a criterion for deciding what kind of control input to perform. For example, the state of the waste incineration plant when a certain control input is performed is predicted, and the state of the waste incineration plant when another control input is performed is predicted, and which control input is more preferable is compared. In addition, it becomes possible to determine the content of the control input that is actually performed.

As described above, the in-furnace image used for prediction may be an image showing the state of the garbage incineration process in chronological order. , the following configuration may be adopted.

That is, the image preprocessing unit 101 transmits to the encoder 103 a plurality of in-furnace images captured in one movement period of the subject among a plurality of time-series in-furnace images in which a subject moving at a predetermined cycle is captured. may be entered. If the subject that moves at a predetermined cycle is a grate, the predetermined cycle is defined as the time from when the grate reaches one end of its range of motion to when it reaches the other end. good.

According to the above configuration, since a plurality of images captured in one movement period of the subject are input to the encoder 103, the encoder 103 receives a latent variable reflecting how the subject behaves in one movement period. can be output. As a result, it is possible to accurately predict the future state of the refuse incineration plant by properly reflecting changes in the state of the refuse incineration plant for each operation cycle of the subject. In this case, the PD preprocessing unit 102 removes noise from the process data acquired in one operation cycle and inputs the process data to the encoder 103 .

[Network structure of VAE]
A network structure of a VAE (Variational Auto-Encoder) will be described with reference to FIG. FIG. 3 is a diagram showing an example of a VAE network structure. As shown, encoder 103 and decoder 105 constitute a VAE. Note that FIG. 3 omits description of a configuration related to prediction for the sake of simplicity.

Encoder 103 includes CNN 1031 , FCN 1032 and FCN 1033 . Details will be described below, but CNN 1031 and FCN 1032 are for processing in-furnace images, and FCN 1033 is for processing process data. With these configurations, the encoder 103 can output latent variables from the in-furnace image and the process data, which have different properties.

Decoder 105 also includes FCN 1051 , FCN 1052 and CNN 1053 . Details will be described below, but FCN 1051 and CNN 1053 are for processing in-furnace images, and FCN 1052 is for processing process data. As a result, the decoder 105 can decode the latent variables and output data with different properties, such as the in-furnace image and the process data.

CNN1031 is a convolutional neural network (Convolutional Neural Network). The CNN 1031 is for extracting the feature points of the in-core image _xt , receives the in-core image _xt as an input, and outputs feature data indicating the feature points of the in-core image _xt . The feature data is lower-dimensional data than the in-core image _xt , which indicates the feature points of the in-core image _xt .

The role of the encoder 103 is to output the probability distribution of the latent variable _at . Here, the probability distribution of the latent variable _at is assumed to be a Gaussian distribution, that is, a normal distribution, and the mean μ and the covariance matrix Σ are output. Specifically, FCN 1032 outputs mean μ, and FCN 1033 outputs covariance matrix Σ, as described below.

Both FCN 1032 and FCN 1033 are fully-connected networks. The FCN 1032 receives as inputs the feature data of the in-core image _xt output by the CNN 1031 and the process data _yt from which noise has been removed by preprocessing, and outputs the average μ in the probability distribution of the latent variable _at . Also, the FCN 1033 receives the feature points of the in-core image _xt extracted by the CNN 1031 and the process data _yt as inputs, and outputs the covariance matrix Σ in the probability distribution of the latent variable _at .

CNN 1053 provided in decoder 105 is also a convolutional neural network, and FCN 1051 and FCN 1052 are also fully connected networks. The FCN 1051 receives the latent variable _at as an input and outputs feature data of the in-core image. Then, the CNN 1053 receives the feature data of the in-core image and outputs the in-core image. Also, the FCN 1052 receives the latent variable _at as an input and outputs process data.

Such a VAE can be constructed by optimizing each parameter (neural network weight) by pre-learning. For example, each parameter may be determined by learning so as to maximize an evidence lower bound (ELBO). Other parameters in KVAE, which is a combined VAE and LGSSM approach, are similar.

Since the process data _yt with noise removed is input to the encoder 103, the value obtained by decoding the latent variable _at is not the value of the process data itself, but the value of the process data obtained by removing noise from the process data by preprocessing. data value. The same is true for in-furnace images, and the CNN 1053 outputs preprocessed in-furnace images.

Further, as described above, by setting the latent variable a _t+1 predicted by the prediction unit 104 as the latent variable to be decoded, the predicted value of the process data at the previous time (time t+1 in this example) and the furnace An inner image (predicted image) can be output.

[Details of the state space model]
More specifically _, the _state space model used by the prediction unit 104 is a dynamic model p _γt ( _at |u _t ). With this model, the latent variable at+1 at the next time t+ ₁ can be obtained. As described above, in this embodiment, an example using LGSSM as the dynamic model p _γt (at | _{u t} ) will be described.

When LGSSM is used, the relationship between the latent variable a _t and the control input data u _t is expressed as a linear relationship by the state variable z _t and parameters γ _t =[A _t , B _t , C _t ]. Note that _At is a transition matrix that determines transition destinations of state variables (hidden states), _Bt is a control matrix that controls transition destinations according to control input data _ut , and _Ct is an output matrix. These matrices are all time-varying (change over time) matrices. In the prediction using the LGSSM, the prediction unit 104 nonlinearly changes the parameters of the LGSSM using a dynamic parameter network, as described below.

Details of the state space model will be described with reference to FIG. FIG. 4 is a diagram showing a configuration example of a state space model. FIG. 4(a) shows a configuration example of the state space model in the observation step, and FIG. 4(b) shows a configuration example of the state space model in the prediction step. Both state space models in FIGS. 4A and 4B include a dynamic parameter network having LSTM (Long Short-Term Memory) 1041 and FCN 1042 as components. Although the details will be described below, the dynamic parameter network is a network in which a plurality of state space models are linearly combined with weights _αt to form one state space model.

The LSTM 1041 outputs _{an output dt obtained} by non-linearly changing the latent variable at for each time. Then, FCN 1042 outputs γ _t that is a linear combination of K different state-space models using weight α _t calculated by applying the softmax function to the output d _t . The K different state-space models are represented by the K parameters Γ in the LGSSM. Note that the parameter Γ is represented by the following formula (2).

…Formula (1)
Since the weight α _t is calculated from the current latent variable a _t and the past output d _t−1 of the LSTM 1041, the transition matrix A _t is expressed by the following equation (2).

…Formula (2)
Also, the control matrix B _t and the output matrix C _t (the matrix used when outputting the latent variable a _t+1 from the state variable z _t+1 ) are obtained in the same manner.

From the above, the parameter γ _t is represented by the following formula (3). By using a dynamic parameter network with such a configuration, the prediction unit 104 can handle nonlinear state transitions.

…Formula (3)
Here, the transition of the state variable from _zt to zt ₊₁ is represented by the following equation (4).
zt ₊₁ =At* _zt + _Bt * _ut ... _Formula (4)
Therefore, the prediction unit 104 obtains the predicted values of the state variables from the transition matrix A _t and the control matrix B _t calculated as described above.

can be asked for. Then, the prediction unit 104 obtains the predicted value of the latent variable from the predicted value of the state variable by the output matrix _Ct .

can be calculated and output.

That is, in the observation step shown in FIG. 4A, the latent variable at at time _t output by the encoder 103 is input to the dynamic parameter network. This latent variable a _t is generated by the encoder 103 from input data (preprocessed in-furnace image and noise-removed process data) at time t. Then, the LSTM 1041 of the dynamic parameter network outputs _dt , and the FCN 1042 outputs the weight _αt using this _dt as an input. Furthermore, the prediction unit 104 determines the parameter γ _t by the above equation (3), and uses this γ _t to calculate the state variable z _t from the control input data u _t . The state variable z _t is the state variable predicted based on the latent variable a _t-1 at time t-1

used to update the

On the other hand, in the prediction step shown in (b) of FIG.

is the predicted value of the latent variable by the output matrix C _t from

Calculate Subsequently, the prediction unit 104 inputs the predicted values of the latent variables to the dynamic parameter network. As a result, the LSTM 1041 of the dynamic parameter network outputs _dt , and the FCN 1042 outputs the weight _αt using this _dt as an input. Furthermore, the prediction unit 104 determines the parameter γ _t according to the above equation (3), and uses this γ _t to obtain the predicted value of the state variable at time t+1 from the control input data u _t .

Calculate Then, the prediction unit 104 obtains the predicted value of the latent variable from the predicted value of the state variable by the output matrix _Ct .

is calculated and output.

As described above, in the observation step, the prediction unit 104 calculates the state variables based on the state variables generated by the encoder 103 from the latest input data, and uses the calculated state variables to predict the previously predicted state variables. Update predictions. Then, in the prediction step, the prediction unit 104 calculates predicted values of the state variables from the updated state variables and the control input data, and calculates and outputs predicted values of the latent variables from the predicted values of the state variables. By repeating these steps, the prediction unit 104 can output a predicted value of a latent variable with high prediction accuracy.

[Process flow]
The flow of processing executed by the information processing apparatus 1 will be described with reference to FIG. FIG. 5 is a flowchart showing an example of processing (information processing method) executed by the information processing apparatus 1 .

In S1, the image preprocessing unit 101 receives input of the in-furnace image, the PD preprocessing unit 102 receives input of process data, and the prediction unit 104 receives input of control input data _ut . Then, in S2, the image preprocessing unit 101 preprocesses the in-furnace image to generate a preprocessed in-furnace image _xt , and the PD preprocessing unit 102 preprocesses the process data to remove noise. of process data _yt .

In S3, the encoder 103 extracts feature points of the in-furnace image _xt using the CNN 1031 to generate feature data. In subsequent S4, the encoder 103 connects the feature data generated in S3 and the process data _yt by the FCN 1032 and outputs the parameter μ. Also, in S5, the FCN 1033 of the encoder 103 connects the feature data generated in S3 and the process data _yt by the FCN 1032 to output the covariance matrix Σ. Then, in S6 (a step of outputting a latent variable), the encoder 103 outputs a latent variable at from the parameter μ output in S4 and the covariance matrix _Σ output in S5.

In S7, the prediction unit 104 uses the dynamic model to calculate the latent variable _at calculated in S6 and the control input _data ut received in S1 to obtain the future value of the latent variable _at . Calculate the variable a _t+1 (prediction step). Then, in S8 (a step of outputting a predicted value), the FCN 1051 and CNN 1053 of the decoder 105 decode the latent variable at ₊₁ to generate and output an in-core image xt ₊₁ , which is a predicted image. In S9 (a step of outputting a predicted value), the FCN 1052 of the decoder 105 decodes the latent variable at ₊₁ to calculate and output the predicted value yt ₊₁ of the process data (more precisely, the noise-removed process data). . The processes of S8 and S9 may be performed simultaneously, or the process of S9 may be performed first.

[Embodiment 2]
Other embodiments of the invention are described below. For convenience of description, members having the same functions as those of the members described in the above embodiments are denoted by the same reference numerals, and description thereof will not be repeated. The same applies to the third and subsequent embodiments.

FIG. 6 is a diagram showing an outline of prediction by the information processing device 1a of this embodiment. The information processing apparatus 1a differs from the information processing apparatus 1 shown in FIG. 1 in that the prediction section 104 is replaced with a prediction section 104a. Like the prediction unit 104, the prediction unit 104a predicts the future value of the latent variable from the value of the latent variable output by the encoder 103 using a dynamic model that models the variation of the latent variable. be.

The prediction unit 104a performs prediction using LSTM, as shown in FIG. In other words, the information processing device 1a performs prediction using the VAE-LSTM technique. Specifically, in the example of FIG _. 6, the prediction unit _104a uses the LSTM to obtain a variable d Predict _t+1 . Then, in the prediction step, prediction section 104 converts variable d _t+1 into future latent variable a _{t+1 by LSTM and outputs the future latent variable a t+1} to decoder 105 .

[Embodiment 3]
In the second embodiment, prediction is performed using LSTM, which is a type of RNN (Recurrent Neural Network), but prediction may be performed using another type of RNN. In this case, the prediction unit 104 uses another type of RNN to predict the variable dt ₊₁ that indicates the future state of the waste incineration plant from the latent variable at output from the encoder ₁₀₃ and the control input data _ut . . Then, in the prediction step, the prediction unit 104 converts the variable d _t+1 into a future latent variable a t+1 using the other type of RNN, and outputs the future latent variable a _t+1 to the decoder 105 .

In addition, in Embodiments 1 and 2, VAE is used for encoding and decoding, but instead of VAE, AE (Auto-Encoder) that does not consider probability distribution may be used for encoding and decoding. Even when AE is used, the predicted values of the predicted image and the process data can be calculated by the processing shown in FIG. In addition, various techniques can be applied to prediction using AE as in the case of using VAE. For example, it is possible to calculate predicted values of predicted images and process data in the framework of the Kalman filter (using dynamic model learning of real vectors and a predictor) as in the first embodiment. Similarly, RNN can be used to calculate predicted values of predicted images and process data as in the second embodiment. Furthermore, as described above in this embodiment, RNN can also be used to calculate predicted values for predicted images and process data.

[Embodiment 4]
The information processing device 1 described above can be incorporated into a plant management system as shown in FIG. 7, for example. FIG. 7 is a block diagram showing an example of a plant management system 5 including the information processing device 1. As shown in FIG. The plant management system 5 includes a sensor 51 , a control input recording device 52 , an imaging device 53 , a plant information storage device 54 , a display device 55 and a plant control device 56 in addition to the information processing device 1 . It is also possible to incorporate the information processing device 1a instead of the information processing device 1. FIG.

The sensor 51 is installed in the plant, detects the magnitude and amount of change in a predetermined physical quantity (e.g., temperature, pressure, flow rate, light, magnetism, etc.) relating to a predetermined process in the plant, and outputs the results in chronological order. do. The output value of the sensor 51 constitutes part of the process data used by the information processing device 1 .

The control input recording device 52 detects that a control input (for example, a control input by a manual operation by an operator) has been performed in the plant, and stores history information of the control input in the plant information storage device 54 . The control input history information may include the content of the executed control input and the date and time when the control input was performed. The information processing device 1 can use this history information as process data.

The photographing device 53 photographs how a predetermined process proceeds in the plant. For example, if the predetermined process is a garbage incineration process, the photographing device 53 photographs the inside of the incinerator to generate an in-furnace image.

The plant information storage device 54 stores various data acquired while a predetermined process is in progress in the plant. Specifically, the output value of the sensor 51, the history information recorded by the control input recording device 52 and the plant control device 56, and the image captured by the imaging device 53 are stored.

The display device 55 is a device that displays an image. For example, the display device 55 can display prediction results (predicted images and predicted values of process data) output by the information processing device 1 . The display device 55 may be placed in a plant control room or the like so that a worker who performs control input to the plant can confirm the above prediction results.

The plant controller 56 controls operations of various devices included in the plant. For example, in the case of a waste treatment plant, the plant control device 56 controls the feeding of waste into the incinerator, the speed of transporting waste within the incinerator, the flow rate of air fed into the incinerator, and the like. Then, the plant control device 56 causes the plant information storage device 54 to store history information of the executed control (control content, execution date and time, etc.). The information processing device 1 can use this history information as process data.

In the plant management system 5, the sensor 51, the control input recording device 52, and the plant control device 56 store various data in the plant information storage device 54 as described above. In addition, the photographing device 53 causes the plant information storage device 54 to store an image showing how the process progresses. Then, the information processing device 1 acquires the data stored in the plant information storage device 54 as process data and images for prediction, and uses the acquired data to predict the future state of the plant.

The information processing device 1 that has predicted the future state of the plant may cause the display device 55 to display the prediction result as post-processing of the prediction. This allows plant personnel to be made aware of future plant conditions and to make necessary control inputs. In addition, the information processing device 1 determines whether or not the predicted future state of the plant requires a control input by the operator. may be displayed. Further, in this case, the information processing device 1 may display the content of the control input to be executed.

Further, the information processing device 1 may control the plant via the plant control device 56 based on the prediction result. For example, if the prediction result indicates that the temperature inside the incinerator will drop, the information processing device 1 performs control to prevent the temperature inside the incinerator from dropping (for example, the amount of garbage supplied to the incinerator or control to increase the amount of air supplied) may be caused to be performed by the plant control device 56 .

The content of the control input corresponding to the prediction result may be determined using a learned model obtained by modeling the relationship between the state of the plant and the control input to be executed in the state of the plant by machine learning. At least one of the process data and the predicted image may be used as the information indicating the state of the plant for the teacher data used in the machine learning.

[Example of realization by software]
The control blocks of the information processing apparatus 1 (especially each block included in the control unit 10 in FIG. 1) may be realized by a logic circuit (hardware) formed in an integrated circuit (IC chip) or the like, or by software. may be realized.

In the latter case, the information processing apparatus 1 is provided with a computer that executes instructions of a program, which is software that implements each function. This computer includes, for example, one or more processors, and a computer-readable recording medium storing the program. In the computer, the processor reads the program (information processing program) from the recording medium and executes it, thereby achieving the object of the present invention. As the processor, for example, a CPU (Central Processing Unit) can be used. As the recording medium, a "non-temporary tangible medium" such as a ROM (Read Only Memory), a tape, a disk, a card, a semiconductor memory, a programmable logic circuit, or the like can be used. In addition, a RAM (Random Access Memory) for developing the above program may be further provided. Also, the program may be supplied to the computer via any transmission medium (communication network, broadcast wave, etc.) capable of transmitting the program. Note that one aspect of the present invention can also be implemented in the form of a data signal embedded in a carrier wave in which the program is embodied by electronic transmission.

[Modification]
The executing subject of each process described in each of the above embodiments can be changed as appropriate. For example, any one or more of S1 to S9 in FIG. 5 may be executed by another device (for example, a cloud server) that can communicate with the information processing device 1. FIG. In other words, in the information processing method described with the information processing apparatus 1 as the execution subject, the execution subject of the following steps may be one information processing apparatus 1 or a plurality of different information processing apparatuses.

Encoding time-series data for a given process in the plant to output latent variables. A prediction step of predicting a future value of the latent variable from the value of the latent variable output by the encoder using a dynamic model that models fluctuations in the latent variable. decoding the value of the latent variable predicted in the prediction step and outputting the predicted value of the time-series data;

The present invention is not limited to the above-described embodiments, but can be modified in various ways within the scope of the claims, and can be obtained by appropriately combining technical means disclosed in different embodiments. is also included in the technical scope of the present invention.

[Reference example]
As in the example of FIG. 3, even if the latent variables output by the encoder 103 are input to the decoder 105 without prediction and decoded, the data before encoding and the data after decoding are different. may not match. Of course, even when prediction is made, there may be a discrepancy between the predicted data and the actually measured data. Also, by gradually changing the parameters of the latent variables, it is possible to output slightly different data to the decoder 105 .

Using this point, the encoder 103 and decoder 105 (and the prediction unit 104 if necessary) can automatically generate teacher data for machine learning. This makes it possible to make up for the lack of teacher data necessary for machine learning to achieve the required determination accuracy.

For example, when generating teacher data for machine learning the correlation between the in-furnace image and the process data, the actually measured process data and the in-furnace image are input to the encoder 103, and the in-furnace image output from the decoder 105 is generated. The image may be associated with the actually measured process data and used as teacher data. Then, by gradually changing the parameters of the latent variables, the decoder 105 is caused to output slightly different in-furnace images, and each output in-furnace image is associated with the actually measured process data as teacher data. good too.

1, 1a information processing device 103 encoders 104, 104a prediction unit 105 decoder

Claims (6)

an encoder that encodes time-series data for a given process in the plant and outputs a latent variable;
a prediction unit that predicts a future value of the latent variable from the value of the latent variable output by the encoder using a dynamic model that models fluctuations in the latent variable;
a decoder that decodes the value of the latent variable predicted by the prediction unit and outputs a predicted value of the time-series data ;
The time-series data includes an image of the progress of the process and process data acquired during the progress of the process,
The encoder outputs the latent variable reflecting the feature points of the image and the feature points of the process data,
wherein the decoder decodes the value of the latent variable predicted by the prediction unit, and outputs a predicted value of the process data and a predicted image showing a future state of the process. Device. an encoder that encodes time-series data for a given process in the plant and outputs a latent variable;
a prediction unit that predicts a future value of the latent variable from the value of the latent variable output by the encoder using a dynamic model that models fluctuations in the latent variable;
a decoder that decodes the value of the latent variable predicted by the prediction unit and outputs a predicted value of the time-series data;
The dynamic model is a model of the correlation between the time-series data, the control input performed at the plant, and the variation of the latent variable,
The prediction unit uses the control input data indicating the content of the control input to the plant, and calculates, from the value of the latent variable output by the encoder, An information processing device characterized by predicting a value of a latent variable. inputting to the encoder a plurality of images captured in one movement cycle of the subject, from among a plurality of time-series images photographing the progress of the process, in which the subject moving at a predetermined cycle is captured; The information processing apparatus according to claim 1 , characterized by: An information processing method by an information processing device,
encoding time series data for a given process in the plant to output latent variables;
a prediction step of predicting a future value of the latent variable from the value of the latent variable output by encoding using a dynamic model that models fluctuations in the latent variable;
decoding the value of the latent variable predicted in the prediction step and outputting the predicted value of the time series data ;
The time-series data includes an image of the progress of the process and process data acquired during the progress of the process,
In the step of outputting the latent variables, outputting the latent variables reflecting the feature points of the image and the feature points of the process data,
In the step of outputting the predicted value, the value of the latent variable predicted in the prediction step is decoded to output the predicted value of the process data and a predicted image showing the future state of the process. An information processing method characterized by: An information processing method by an information processing device,
encoding time series data for a given process in the plant to output latent variables;
a prediction step of predicting a future value of the latent variable from the value of the latent variable output by encoding using a dynamic model that models fluctuations in the latent variable;
decoding the value of the latent variable predicted in the prediction step and outputting the predicted value of the time series data;
The dynamic model is a model of the correlation between the time-series data, the control input performed in the plant, and the fluctuation of the latent variable,
In the predicting step, using control input data indicating the content of the control input to the plant, from the value of the latent variable output in the step of outputting the latent variable, the control input having the content indicated by the control input data. An information processing method characterized by predicting the value of the latent variable after the is performed . 3. An information processing program for causing a computer to function as the information processing apparatus according to claim 1 or 2 , the information processing program for causing a computer to function as the encoder, the prediction unit, and the decoder.

Images