Information processing device and information processing method

Abstract

It accurately and instantly estimates the conditions inside the incinerator. [Solution] An information processing device comprising a control unit that acquires input parameters including measurement results of physical quantities inside an incinerator and estimates the state inside the incinerator using a neural network learning model based on the input parameters, wherein the control unit generates a learning model by adjusting the parameters of the intermediate layer of the neural network based on the calculation result of a loss function that incorporates differential values ​​showing the physical relationship between the input parameters and the output parameters of the neural network, inputs the input parameters into the generated learning model, and outputs the output parameters of the neural network of the learning model as the estimated state inside the incinerator.

Description

[Technical Field]

[0001] The present invention relates to an information processing apparatus and an information processing method. [Background technology]

[0002] One example of an invention for predicting the state inside an incinerator is the invention disclosed in Patent Document 1. The learning model generation method disclosed in Patent Document 1 is a method for generating a learning model that takes the input amount of fuel (city gas and air), which is a control variable, and furnace state information as input, and learns to output a predicted value of the target variable corresponding to the state information. The learning model generated by this method, for example, when the current values ​​of the input amount of fuel gas and state information (such as the input amount of waste liquid) are input, outputs a predicted value of the exhaust gas concentration in the combustion furnace. [Prior art documents] [Patent Documents]

[0003] [Patent Document 1] Patent No. 7250205 [Overview of the project] [Problems that the invention aims to solve]

[0004] In the case of the learning model disclosed in Patent Document 1, while it outputs predicted values ​​through machine learning, the reliability of the prediction results is lower than that of the interpolation region because the extrapolation region is not trained. Furthermore, while there are methods to predict the state inside an incinerator using numerical simulations, these methods are computationally intensive, making it difficult to predict the state inside the incinerator immediately.

[0005] The present invention has been made in view of the above, and aims to estimate the state inside an incinerator accurately and immediately. [Means for solving the problem]

[0006] An information processing device according to one aspect of the present invention is an information processing device comprising a control unit that acquires input parameters including measurement results of physical quantities inside an incinerator and estimates the state inside the incinerator using a neural network learning model based on the input parameters, wherein the control unit generates a learning model by adjusting the parameters of the intermediate layer of the neural network based on the calculation result of a loss function that incorporates differential values ​​showing the physical relationship between the input parameters and the output parameters of the neural network, inputs the input parameters into the generated learning model, and outputs the output parameters of the neural network of the learning model as the estimated state inside the incinerator.

[0007] Furthermore, in an information processing apparatus according to one aspect of the present invention, the loss function may include a time component.

[0008] Furthermore, an information processing device according to one aspect of the present invention stores a plurality of learning models, acquires input parameters, acquires a learning model from the plurality of stored learning models that corresponds to the acquired input parameters, generates a learning model by adjusting the parameters of the intermediate layer of the neural network of the learning model based on the calculation result of a loss function that incorporates the differential value showing the physical relationship between the acquired input parameters and the output parameters of the neural network of the acquired learning model, inputs the input parameters into the generated learning model, and outputs the output parameters of the neural network of the learning model as the estimated state inside the incinerator.

[0009] An information processing method according to one aspect of the present invention is an information processing method executed by an information processing device equipped with a control unit that acquires input parameters including measurement results of physical quantities inside an incinerator and estimates the state inside the incinerator using a neural network learning model based on the input parameters, wherein the device generates a learning model by adjusting the parameters of the intermediate layer of the neural network based on the calculation result of a loss function that incorporates differential values ​​showing the physical relationship between the input parameters and the output parameters of the neural network, inputs the input parameters into the generated learning model, and outputs the output parameters of the neural network of the learning model as the estimated state inside the incinerator. [Effects of the Invention]

[0010] According to the present invention, the state inside the incinerator can be estimated accurately and immediately. [Brief explanation of the drawing]

[0011] [Figure 1] Figure 1 shows the configuration of a waste incineration system according to an embodiment. [Figure 2] Figure 2 is a block diagram showing the configuration of the estimation device according to this embodiment. [Figure 3] Figure 3 is a block diagram showing the configuration of the control unit according to an embodiment. [Figure 4] Figure 4 is a functional block diagram of the control unit. [Figure 5] Figure 5 shows the process flow for retraining a learned model. [Modes for carrying out the invention]

[0012] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. Note that the present invention is not limited by the embodiments described below. Also, in the description of the drawings, the same or corresponding elements are appropriately assigned the same reference numerals. Furthermore, it should be noted that the drawings are schematic, and the dimensional relationships of each element may be different from the actual ones. There may also be parts where the dimensional relationships and ratios between the drawings are different from each other.

[0013] FIG. 1 is a diagram showing the configuration of a waste incineration system 1000 according to an embodiment of the present invention. The waste incinerator 1 is, for example, a grate-type incinerator and includes a combustion chamber 2 and an inlet 3. The inlet 3 is, for example, an inlet for introducing waste W such as industrial waste or household garbage into the combustion chamber 2, and is provided above the combustion chamber 2 on the upstream side of the flow of waste W in the combustion chamber 2. Below the inlet 3, an extruder (not shown) for pushing the introduced waste W into the combustion chamber 2 is arranged, and the waste W introduced into the inlet 3 is pushed into the combustion chamber 2 by the extruder.

[0014] At the bottom of the combustion chamber 2, a grate 5 for burning while moving the waste W is provided. The grate 5 is composed of a drying grate 5a, a combustion grate 5b, and a post-combustion grate 5c, and is arranged in the order of the drying grate 5a, the combustion grate 5b, and the post-combustion grate 5c in the moving direction of the waste W from the inlet 3 side. On the drying grate 5a, mainly the drying, ignition, and initial combustion of the waste W are carried out. On the combustion grate 5b, mainly the thermal decomposition and partial oxidation of the waste W are carried out. Also, on the combustion grate 5b, the combustion of the pyrolysis gas containing carbon monoxide, hydrocarbons, etc. generated by thermal decomposition and the solid content is carried out. On the post-combustion grate 5c, post-combustion is carried out to completely burn the unburned portion of the waste W that has remained unburned. By this post-combustion, a layer of incineration ash after complete combustion is formed on the post-combustion grate 5c.

[0015] Above the downstream side along the flow direction of the waste W in the combustion chamber 2, a boiler 4 is connected. Near the inlet of the boiler 4, a secondary combustion chamber 11 that burns unburned gas in the gas discharged from the combustion chamber 2 is formed. Secondary combustion air is blown into the secondary combustion chamber 11 by a nozzle (not shown). In the secondary combustion chamber 11, the unburned components in the combustion gas generated in the combustion chamber 2 are secondarily combusted by the secondary combustion air, and the exhaust gas after the secondary combustion is heat-recovered by the boiler 4.

[0016] The boiler 4 that performs heat recovery from the exhaust gas has two bending portions 12 and 13 that bend the flow path of the exhaust gas. By these bending portions 12 and 13, a first radiation chamber 14, a second radiation chamber 15, and a convection heat transfer chamber 16 are formed from the upstream side along the flow direction of the exhaust gas. The first radiation chamber 14 through which the exhaust gas flows from the combustion chamber 2 has the secondary combustion chamber 11 as its upstream portion along the flow direction of the exhaust gas. The first radiation chamber 14 and the second radiation chamber 15 are continuous through the bending portion 12, and the lower part of the second radiation chamber 15 and the lower part of the convection heat transfer chamber 16 are continuous through the bending portion 13. The upper end of the convection heat transfer chamber 16 is connected to a dust removal device 23 composed of a bag filter or the like through a flue 21.

[0017] The boiler 4 has an inner wall composed of a refractory wall. The first radiation chamber 14 and the second radiation chamber 15 are provided in a state where heat transfer tubes (not shown) formed by pipes through which steam flows are densely arranged outside the refractory wall forming the inner wall. The heat transfer tubes arranged outside the refractory wall through which water flows become a radiation heat transfer surface that receives radiant heat from the exhaust gas to generate steam and function as an evaporator.

[0018] The convection heat transfer chamber 16 has heat transfer tubes (not shown) arranged in a flag shape at the uppermost part of the flow direction of the exhaust gas. By cooling the exhaust gas flowing into the convection heat transfer chamber 16 with the heat transfer tubes, gaseous or mist-like dust components are solidified and separated from the exhaust gas as dust. The convection heat transfer chamber 16 also includes three superheaters 16A and an economizer 16B, arranged from the upstream side along the flow direction of the exhaust gas. Each superheater 16A has a group of heat transfer tubes arranged in multiple stages in the height direction, with multiple heat transfer tubes arranged horizontally, and the group of heat transfer tubes functions as a convection heat transfer surface. The superheater 16A further superheats the steam generated in the first radiating chamber 14 and the second radiating chamber 15 through heat exchange with the exhaust gas to produce high-temperature, high-pressure superheated steam.

[0019] The economizer 16B is located downstream of the superheater 16A in the direction of exhaust gas flow, and is equipped with heat transfer tubes (not shown). The steam generated in the boiler 4 and used to drive the steam turbine (not shown) is condensed in a condenser (not shown) and flows through the heat transfer tubes of the economizer 16B. The condensate flowing through the heat transfer tubes of the economizer 16B is heated by the residual heat in the exhaust gas after the steam has been superheated by the superheater 16A, and the heated water is supplied to the heat transfer tubes of the first radiating chamber 14 and the second radiating chamber 15, which function as evaporators. The economizer 16B may be installed outside the boiler 4 downstream of the boiler 4 in the direction of exhaust gas flow, rather than inside the convection heat transfer chamber 16, or both an economizer inside the boiler 4 and an economizer outside the boiler 4 may be provided.

[0020] The exhaust gas, whose heat has been recovered by the boiler 4, flows to the cooling tower 17. The cooling tower 17 sprays water into the exhaust gas, lowering its temperature to below 200°C. The exhaust gas, whose temperature has been lowered in the cooling tower 17, flows through the flue 21 to a dust removal device 23, which consists of, for example, a bag filter. In the flue 21, chemicals such as slaked lime and activated carbon are blown into the exhaust gas from a chemical supply device 22. When chemicals are blown into the exhaust gas, they bind to pollutants such as hydrogen chloride and sulfur oxides contained in the exhaust gas. The dust removal device 23 collects and removes the dust and chemicals bound to the pollutants contained in the exhaust gas that has flowed through the flue 21. An induced draft fan 24 is connected to the dust removal device 23. The induced draft fan 24 draws in the exhaust gas from which dust has been removed from the dust removal device 23. The exhaust gas drawn in from the dust removal device 23 by the induced draft fan 24 is sent to the chimney 25 and released into the atmosphere.

[0021] The wind boxes 7a to 7c are located at the bottom of the combustion chamber 2. Specifically, wind box 7a is located below the drying grate 5a, wind box 7b is located below the combustion grate 5b, and wind box 7c is located below the post-combustion grate 5c. In addition, a supply line 31a for supplying primary air used for the combustion of waste W is located below wind box 7a, a supply line 31b for supplying primary air is located below wind box 7b, and a supply line 31c for supplying primary air is located below wind box 7c. Wind box 7a supplies primary air supplied by supply line 31a to the drying grate 5a, wind box 7b supplies primary air supplied by supply line 31b to the combustion grate 5b, and wind box 7c supplies primary air supplied by supply line 31c to the post-combustion grate 5c.

[0022] The supply line 31a is equipped with a damper 32a to adjust the amount of primary air supplied to the windbox 7a, and the supply line 31b is equipped with a damper 32b to adjust the amount of primary air supplied to the windbox 7b. The blower 34 is a device that supplies primary air to the supply lines 31a, 31b, and 31c. The heating device 33 is a device that heats the primary air discharged from the blower 34. The heating device 33 is controlled by a control device (not shown) to control the temperature of the primary air. The damper 32c adjusts the amount of primary air supplied to the supply lines 31a, 31b, and 31c. The dampers 32a, 32b, and 32c are controlled by a control device (not shown) that allows for independent adjustment of the amount of primary air flowing through each of them.

[0023] The primary air supplied from the blower 34 is branched through the heating device 33, the supply pipe 31 that supplies the primary air, and the damper 32c. A portion of the branched primary air is supplied to the drying grate 5a via the damper 32a and the supply line 31a, and a portion of the branched primary air is supplied to the combustion grate 5b via the damper 32b and the supply line 31b. The remaining branched primary air is supplied to the post-combustion grate 5c via the supply line 31c. The primary air supplied from below the grate 5 is supplied to the combustion chamber 2 and used for drying, stirring, and burning the waste W, and also for cooling the grate 5.

[0024] The incinerated ash obtained from combustion in the post-combustion grate 5c is sent from the post-combustion grate 5c to the discharge section 6 located downstream of the waste W flow, and discharged from the discharge section 6 to the outside of the waste incinerator 1.

[0025] Sensor 18 is a group of sensors that detect the conditions inside the combustion chamber 2. For example, sensor 18 detects the pressure inside the combustion chamber 2, the temperature inside the combustion chamber 2, the flow velocity of the gas flowing inside the combustion chamber 2, and the mass fraction of chemical substances inside the combustion chamber 2.

[0026] Figure 2 is a block diagram showing the configuration of the estimation device 100. The estimation device 100 is a device that estimates the state inside the combustion chamber 2 using a neural network, and comprises a control unit 101, an operation unit 102, a display unit 103, an interface 104, and a storage unit 105. The operation unit 102 has a keyboard, mouse, and various buttons for operating the estimation device 100, and is operated by an operator of the waste incineration system 1000. The display unit 103 is, for example, a display device that displays information representing the state inside the combustion chamber 2 estimated by the estimation device 100. The interface 104 acquires the measurement results of the sensor 18.

[0027] The control unit 101 comprises an arithmetic unit and a memory unit. The arithmetic unit is composed of, for example, a CPU (Central Processing Unit). The memory unit comprises, for example, a part composed of ROM (Read Only Memory) and a part composed of RAM (Random Access Memory). The part composed of ROM stores various programs and data used by the arithmetic unit to perform arithmetic processing. RAM is used to store the workspace for the arithmetic unit when it performs arithmetic processing and the results of the arithmetic processing performed by the arithmetic unit.

[0028] The memory unit 105 is made of a non-volatile storage medium. The memory unit 105 stores an operating system (not shown), a learning program 105a for realizing a neural network, a learning model 105b which is the result of machine learning by the neural network, and various data.

[0029] The control unit 101 realizes the function of estimating the state inside the combustion chamber 2 by having the CPU execute a program stored in the memory unit 105. The function of the control unit 101 is realized as a functional unit by having the CPU read and execute a program from the memory unit 105. A block diagram of the functions realized by the control unit 101 is shown in Figure 3.

[0030] The acquisition unit 101a acquires the measurement results from the sensor 18. The learning unit 101b uses the input and output parameters acquired by the estimation device 100 as training data to perform machine learning using a neural network, and writes the learned results as a learning model 105b to the storage unit 105 for storage. When the learning unit 101b generates the learning model 105b using a neural network, learning input parameters and learning output parameters are used. The learning unit 101b may, separately from the neural network performing the learning, store the latest learning model at a predetermined timing in the storage unit 105. When storing the learning model 105b in the storage unit 105, it may be an update where the old learning model 105b is deleted and the latest learning model 105b is stored, or it may be an accumulation where part or all of the old learning model 105b is saved while the latest learning model 105b is stored. The estimation unit 101c derives output parameters 212 from input parameters 211 based on the learning model 105b stored in the memory unit 105, and estimates the state inside the combustion chamber 2.

[0031] Figure 4 is a schematic diagram showing the configuration of the neural network 200 implemented in the control unit 101. The neural network 200 has an input layer 201, an intermediate layer 202, and an output layer 203. The input parameters 211 include the measurement results of the sensor 18, for example, the coordinate in the furnace width direction within the combustion chamber 2 (x), the coordinate in the furnace width direction within the combustion chamber 2 (y), the temperature within the combustion chamber 2 (T), the pressure within the combustion chamber 2 (P), the X component of the flow velocity within the combustion chamber 2 (u), the Y component of the flow velocity within the combustion chamber 2 (v), and the mass fraction of methane gas within the combustion chamber 2 (Y CH4 ), the mass fraction of oxygen (Y O2 ), the mass fraction of water (Y H2O ), the mass fraction of nitrogen (Y N2) It is. The input parameter 211 is input to the input layer 201. The input layer 201 consists of a plurality of nodes, and different parameters are input to each node. For example, u and v are the gas flow velocities measured by flow velocity sensors provided at the furnace inlet where waste W is supplied in the combustion chamber 2, the furnace wall surface, etc. T is the gas temperature measured by a temperature sensor or a flame transmission camera composed of an infrared camera provided at the furnace inlet, the furnace wall surface, the furnace outlet, etc., and P is the gas pressure measured by a pressure sensor provided at the furnace inlet, the furnace wall surface, the furnace outlet, etc. Also, Y CH4 、Y O2 、Y H2O 、Y N2 is the mass fraction of the gas measured by a gas concentration sensor provided at the furnace inlet, the furnace wall surface, the furnace outlet, etc.

[0032] The intermediate layer 202 receives the output from the input layer 201. The intermediate layer 202 has a multi-layer structure including a layer composed of a plurality of nodes that receive the output of the input layer 201.

[0033] The output layer 203 receives the output from the intermediate layer 202 and outputs the output parameter 212 which is the estimation result. The output parameter 212 is, for example, the X-direction component (u_pred) of the estimated flow velocity in the combustion chamber 2, the Y-direction component (v_pred) of the estimated flow velocity in the combustion chamber 2, the estimated temperature (T_pred) in the combustion chamber 2, the estimated pressure (P_pred) in the combustion chamber 2, the estimated mass fraction of methane gas (Y CH4 _pred) in the combustion chamber 2, the estimated mass fraction of oxygen (Y O2 _pred), the estimated mass fraction of water (Y H2O _pred), the estimated mass fraction of nitrogen (Y N2 _pred) in the combustion chamber 2. These output parameters are the overall values including the desired points in the combustion chamber 2.

[0034] The learning unit 101b adjusts the parameters of the intermediate layer 202 so that the calculation results of the loss functions L0 to L9 are small. The loss functions are functions derived from the fluid mass conservation equation, the fluid momentum equation, the fluid energy conservation equation, the transport equation derived from the chemical species conservation equation, the chemical species reaction rate equation, the chemical species diffusion equation, etc., and are functions that incorporate differential values ​​that show the physical relationship between the input parameters, which are physical quantities, and the output parameters, which are physical quantities. In this embodiment, the loss functions L1 to L9 are functions that incorporate the differential values ​​213 of the input parameter 211 and the output parameter 212 so as to satisfy the physical equations that represent the state inside the combustion chamber 2, for example, equations (1) to (9) shown below. By incorporating the differential value 213 into the loss functions so as to satisfy the physical equations that represent the state inside the combustion chamber 2, the generated learning model 105b becomes a model in which the estimation results take physical laws into consideration. In addition to physical equations, it is also possible to incorporate the difference with values ​​such as measurement points inside the combustion chamber 2 and points on boundary conditions into the loss functions L1 to L9. For example, the difference between the estimated temperature inside the combustion chamber 2 (T_pred) and the fluid temperature (T) measured by the sensor 18 may be incorporated as the loss function, or the difference between the estimated pressure (P_pred) and the pressure (P) measured by the sensor 18 may be incorporated as the loss function.

[0035] The loss function L1 in equation (1) is a continuity equation derived from the law of conservation of mass. The loss function L2 in equation (2) and the loss function L3 in equation (3) are momentum conservation equations when the width direction of the combustion chamber 2 is the X direction and the length direction of the combustion chamber 2 is the Y direction. The loss function L4 in equation (4) is an energy conservation equation. The loss function L5 in equation (5) is the transport equation for the mass fraction of methane gas (CH4) in combustion chamber 2. The loss function L6 in equation (6) is the transport equation for the mass fraction of oxygen (O2) in combustion chamber 2. The loss function L7 in equation (7) is the transport equation for the mass fraction of water (H2O) in combustion chamber 2. The loss function L8 in equation (8) is the transport equation for the mass fraction of carbon dioxide (CO2) in combustion chamber 2. The loss function L9 in equation (9) is the transport equation for the mass fraction of nitrogen (N2) in combustion chamber 2. The loss function L0 in equation (10) is the difference between the output parameter 212 and the measured value 214. For example, equation (1) incorporates the velocity u in the x-direction, the velocity v in the y-direction, and the temperature T of the fluid in combustion chamber 2, as well as the derivatives of velocity u, velocity v, and temperature T. The resulting learning model takes into account the law of conservation of mass when estimating the contents of combustion chamber 2.

[0036] In the loss functions L1 to L9, x is the coordinate in the furnace width direction within the combustion chamber 2, y is the coordinate in the furnace width direction within the combustion chamber 2, u is the X component of the flow velocity within the combustion chamber 2, and v is the Y component of the flow velocity within the combustion chamber 2. Also, in the loss functions L1 to L9, p is the pressure within the combustion chamber 2, ρ is the density of the fluid within the combustion chamber 2, μ is the kinematic viscosity coefficient of the fluid within the combustion chamber 2, T is the temperature within the combustion chamber 2, and θ is a dimensionless quantity of temperature. Also, in the loss functions L1 to L9, Cp is the specific heat at constant pressure, Cμ is the coefficient of the temperature gradient effect, K is the thermal conductivity, and D is the diffusion coefficient. Also, in the loss functions L1 to L9, Y CH4 Y is the mass fraction of methane gas. O2 This is the mass fraction of oxygen, Y H2O Y is the mass fraction of water. N2 ω is the mass fraction of nitrogen, and ω is the rate of creation / annihilation of each of these chemical species.

[0037]

number

[0038]

number

[0039]

number

[0040]

number

[0041]

number

[0042]

number

[0043]

number

[0044]

number

[0045]

number

[0046] L0 = Output parameter - Measured value ... (10)

[0047] For example, in numerical simulation, the state inside the combustion chamber 2 is calculated sequentially, starting with the parts that can be calculated. This means that the parts to be calculated are calculated in a predetermined order, which takes time. On the other hand, in this embodiment, by using the neural network 200, it becomes possible to calculate the parts to be calculated in parallel and simultaneously, thus reducing the time required to calculate the state.

[0048] In this embodiment, since multiple loss functions are used, the output of the neural network 200 may fall into a local optimum. To avoid falling into a local optimum, the parameters of the hidden layer may be optimized using a two-stage optimization method: first, optimizing the parameters (weights) of the hidden layer with the ADAM (Adaptive Moment Estimation) optimization algorithm to adjust the optimal parameter range, and then adjusting the optimal parameters of the hidden layer with the L-BFGS (Limited-memory BFGS method) algorithm. Alternatively, methods such as steepest descent, stochastic gradient descent, momentum, AdaGrad (Adaptive Gradient Algorithm), and RMSProp may be used to adjust the optimal parameters.

[0049] Physical phenomena can be classified into stationary and transient. Stationary physical phenomena are those in which the reaction or motion is stable and does not change over time. Transient physical phenomena are those in which the reaction or motion is unstable and the state or mode of reaction changes over time. Because the state of transient physical phenomena changes over time, a time component is incorporated into the physical equations that describe the physical phenomenon. For example, when considering the time component in neural network 200, the loss functions L1 to L9 incorporate the time component and become equations (11) to (19) shown below. In the loss functions L1 to L9 of equations (11) to (19), the term related to the time component is the first term containing ∂t.

[0050]

number

[0051]

number

[0052]

number

[0053]

number

[0054]

number

[0055]

number

[0056]

number

[0057]

number

[0058]

number

[0059] Incidentally, since the estimation of the state inside the combustion chamber 2 requires immediacy, when generating the learning model 105b by incorporating a time component into the loss function L1 to L9, it is necessary to shorten the training time of the neural network 200. However, if the boundary conditions of the neural network 200, such as the amount of waste introduced into the combustion chamber 2, the temperature, and the measured values ​​of the sensor 18, change significantly, it may take time to retrain, and the state inside the combustion chamber 2 may not be obtained immediately. Therefore, to solve this problem, the learning model 105b generated by pre-training for multiple boundary conditions may be stored in the memory unit 105, and when the input parameters 211 for estimation change, the learning unit 101b may retrain the neural network 200 based on the learning model 105b that is close to the changed boundary conditions, and the estimation unit 101c may perform estimation using the learning model 105b generated by the retraining.

[0060] Figure 5 shows the process flow for retraining based on a learning model 105b that is close to the boundary conditions from among multiple already trained learning models 105b. The learning unit 101b acquires measurement values ​​214 from the sensor 18 (step S1). Based on the acquired measurement values ​​214, the learning unit 101b selects one learning model 105b from among multiple learning models 105b stored in the storage unit 105 (step S2).

[0061] Here, for example, the learning unit 101b focuses on one parameter among the measurement values ​​214, which consist of multiple parameters, and selects a learning model 105b generated using an input parameter 211 that is close to the value of the parameter of interest. For example, consider a case where the temperature T inside the combustion chamber 2, which is one of the measurement values ​​214, is focused on, and the memory unit 105 stores the results of training the neural network 200 at 100°C intervals from 100°C to 2000°C. In this case, if the temperature acquired from the sensor 18 is 1030°C, the learning unit 101b acquires the learning model 105b trained at 1000°C from the memory unit 105, and retrains the acquired learning model 105b with the acquired measurement value 214. Alternatively, for example, the square of the difference between the input parameters 211 may be taken, and the distance in high-dimensional space may be used to select a trained learning model 105b, or nonlinear clustering may be performed with a machine learning model, and a learning model 105b with similar properties to the output parameters may be selected. The estimation unit 101c acquires the learning model 105b generated by the learning unit 101b through retraining (step S3), and uses the acquired learning model 105b and the measured value 214 as input parameter 211 to estimate the state inside the combustion chamber 2.

[0062] With this configuration, the learning unit 101b performs training using the learning model 105b, which has similar boundary conditions, thus shortening the training time and reducing the time required to estimate the state inside the combustion chamber 2. [Explanation of Symbols]

[0063] 1. Waste Incinerator 2 Combustion chambers 100 Estimator 101 Control Unit 101a Acquisition Department 101b Learning Department 101c Estimation Section 102 Operation section 103 Display section 104 Interface 105 Storage section 200 Neural Networks 201 Input Layer 202 Middle Class 203 Output Layer 211 Input Parameters 212 Output Parameters 213 Differential Value 214 measurements

Claims

1. An information processing device comprising a control unit that acquires input parameters including measurement results of physical quantities inside an incinerator, and estimates the state inside the incinerator using a neural network learning model based on the input parameters, The control unit, A learning model is generated by adjusting the parameters of the intermediate layer of the neural network based on the calculation result of a loss function that incorporates the differential value showing the physical relationship between the input parameters and the output parameters of the neural network. The input parameters are input to the generated learning model, and the output parameters of the neural network of the learning model are output as the estimated state inside the incinerator. Information processing device.

2. The information processing apparatus according to claim 1, wherein the loss function includes a time component.

3. Multiple of the aforementioned learning models are stored, Get the input parameters, From among the multiple learned models stored in memory, the learned model corresponding to the acquired input parameters is selected. Based on the calculation result of a loss function that incorporates the derivative value showing the physical relationship between the acquired input parameters and the output parameters of the neural network of the acquired learning model, a learning model is generated by adjusting the parameters of the intermediate layer of the neural network of the learning model. The input parameters are input to the generated learning model, and the output parameters of the neural network of the learning model are output as the estimated state inside the incinerator. The information processing apparatus according to claim 1.

4. An information processing method performed by an information processing device comprising a control unit that acquires input parameters including measurement results of physical quantities inside an incinerator and estimates the state inside the incinerator using a neural network learning model based on the input parameters, A learning model is generated by adjusting the parameters of the intermediate layer of the neural network based on the calculation result of a loss function that incorporates the differential value showing the physical relationship between the input parameters and the output parameters of the neural network. The input parameters are input to the generated learning model, and the output parameters of the neural network of the learning model are output as the estimated state inside the incinerator. Information processing methods.

Images