The present disclosure relates to a correction apparatus, a prediction apparatus, a method, a program, and a correction model.
A hitherto known system is configured to predict an operational state of a device such as an air conditioner from operational data of the device using a prediction model generated by machine learning, and control operations of the device or diagnose failures of the device (Japanese Patent Application Publication No. 2020-109581).
An apparatus according to an aspect of the present disclosure is an apparatus configured to perform correction regarding a predicted value for an operational state predicted from operational data of a device, and includes:
Embodiments of the present disclosure will be described below with reference to the drawings.
An apparatus configured to perform correction regarding a predicted value for an operational state predicted from operational data of a device (hereinafter, the apparatus may also be referred to as a correction apparatus) will be described below. The correction apparatus includes a provisional prediction model that is trained by machine learning based on training data including operational data and an operational state of a device different from the device, and a correction unit configured to perform correction regarding a predicted value for the operational state of the device predicted using the provisional prediction model.
Specifically, the correction apparatus can generate a correction model configured to correct a predicted value for the operational state of the device predicted using the provisional prediction model (embodiment 1). The correction apparatus can also correct an abnormality determination threshold using the predicted value for the operational state of the device predicted using the provisional prediction model (embodiment 2). The correction apparatus can correct a control gain involved in a control using the predicted value for the operational state of the device predicted using the provisional prediction model (embodiment 3).
<Generation of Provisional Prediction Model>
First, a provisional prediction model generation apparatus 600 generates a provisional prediction model 10. Specifically, the provisional prediction model generation apparatus 600 generates a provisional prediction model (i.e., a model for the device type B) 10 by performing machine learning using training data (specifically, by performing machine learning by associating operational data and an operational state of a device that is device type B with each other).
<Generation of Correction Model>
Next, a correction model generation apparatus (an example of the correction apparatus) 400 generates a correction model 20. The correction model 20 is a model configured to correct a predicted value for an operational state.
Specifically, the correction model generation apparatus 400 acquires a predicted value for an operational state of a device that is device type A, by inputting operational data of the device that is device type A into the provisional prediction model (i.e., the model for the device type B) 10 to output an operational state. The correction model generation apparatus 400 acquires an actually measured value for the operational state of the device that is device type A (specifically, the actually measured value is an operational state calculated from the operational data of the device that is device type A). Then, the correction model generation apparatus 400 generates the correction model 20 by performing machine learning by associating the predicted value for the operational state of the device that is device type A with the actually measured value for the operational state of the device that is device type A (note that the machine learning may be performing by associating the predicted value for the operational state of the device that is device type A and the operational data of the device that is device type A with the actually measured value for the operational state of the device that is device type A). In this case, the operational data of the device type A necessary for generating the correction model is typically not operational data accumulated over a long period of time necessary for generating a prediction model for the device type A, and need only be operational data accumulated over a short period of time. For example, operational data acquired in a test room during development of the device type A, and data of a test operation during installation may be used.
<Operation Using Provisional Prediction Model and Correction Model in Combination>
Subsequently, the device that is device type A starts to be operated. A prediction apparatus 500 predicts an operational state from operational data of the device that is device type A.
Specifically, the prediction apparatus 500 acquires a predicted value for an operational state of the device that is device type A by inputting operational data of the device that is device type A into the provisional prediction model (i.e., the model for the device type B) 10 to output an operational state. Moreover, the prediction apparatus 500 acquires a corrected predicted value for the operational state of the device that is device type A by inputting the predicted value for the operational state of the device that is device type A into the correction model 20 to output the corrected predicted value for the operational state (the corrected predicted value for the operational state of the device that is device type A may be acquired by inputting the predicted value for the operational state of the device that is device type A and the operational data of the device that is device type A into the correction model 20).
In this way, by generating a correction model using a small amount of operational data of the device that is device type A (e.g., a new device type), and using this correction model and the provisional prediction model for the device type B (e.g., an old device type), it is possible to predict an operational state from the operational data of the device that is device type A without generating a prediction model dedicated to the device type A.
<Example of Overall Configuration>
As illustrated in <Example 1>, the prediction apparatus 500 may be implemented on a computer installed in, for example, the same building as that in which an air conditioning system 100 is installed. The correction model generation apparatus 400 may be implemented on a cloud server apart from the air conditioning system 100 and the prediction apparatus 500.
As illustrated in <Example 2>, the prediction apparatus 500 may be implemented as a part of the air conditioning system 100 (e.g., may be installed in an outdoor unit 200 or an indoor unit 300). The correction model generation apparatus 400 may be implemented on a cloud server apart from the air conditioning system 100 and the prediction apparatus 500.
As illustrated in <Example 3>, the correction model generation apparatus 400 and the prediction apparatus 500 may be implemented on a cloud server apart from the air conditioning system 100. The correction model generation apparatus 400 and the prediction apparatus 500 may be implemented on one apparatus.
As illustrated in <Example 4>, the correction model generation apparatus 400 and the prediction apparatus 500 may be implemented on a computer installed in, for example, the same building as that in which the air conditioning system 100 is installed. The correction model generation apparatus 400 and the prediction apparatus 500 may be implemented on one apparatus.
As illustrated in <Example 5>, the correction model generation apparatus 400 and the prediction apparatus 500 may be implemented as a part of the air conditioning system 100 (e.g., may be installed in an outdoor unit 200 or an indoor unit 300). The correction model generation apparatus 400 and the prediction apparatus 500 may be implemented on one apparatus.
The hardware configuration of the air conditioning system 100 will be described with reference to
<Hardware Configuration of Air Conditioning System (in a Case of Air-Cooling Operation)>
In the example of
<<Outdoor Unit>>
In the outdoor unit 200, the outdoor heat exchanger 201, the compressor 202, the supercooling heat exchanger 203, the supercooling heat exchanger expansion valve (bypass circuit) 204, and the outdoor unit main expansion valve (main refrigerant circuit) 205 are connected to piping. The outdoor unit 200 includes various sensors (e.g., temperature sensors (e.g., thermistors) (1), (3), (4), (6), and (7), and pressure sensors (2) and (5)).
<<Indoor Unit>>
In the indoor unit 300, the indoor heat exchanger 301 and the indoor heat exchanger expansion valve 302 are connected to piping. The indoor unit 300 includes various sensors (e.g., temperature sensors (e.g., thermistors) (8) and (9)).
<Hardware Configuration of Air Conditioning System (in a Case of Air-Warming Operation)>
In the example
<<Outdoor Unit>>
In the outdoor unit 200, the outdoor heat exchanger 201, the compressor 202, the supercooling heat exchanger 203, a supercooling heat exchanger expansion valve (bypass circuit) 204, and the outdoor unit main expansion valve (main refrigerant circuit) 205 are connected to the piping. The outdoor unit 200 includes various sensors (e.g., temperature sensors (e.g., thermistors) (1), (3), (4), (6), and (7), and pressure sensors (2) and (5)).
<<Indoor Unit>>
In the indoor unit 300, the indoor heat exchanger 301 and the indoor heat exchanger expansion valve 302 are connected to a pipe. The indoor unit 300 includes various sensors (e.g., temperature sensors (e.g., thermistors) (8) and (9)).
<Hardware Configuration of Air Conditioning System (in a Case of Simultaneous Air-Cooling and Warming Operation)>
The present disclosure is not limited to an ai-cooling operation and an air-warming operation, and can be applied to a simultaneous air-cooling and warming operation. A simultaneous air-cooling and warming operation will be described below with reference to
<Hardware Configurations of Correction Model Generation Apparatus, Prediction Apparatus, and Provisional Prediction Model Generation Apparatus>
The correction model generation apparatus 400, the prediction apparatus 500, and the provisional prediction model generation apparatus 600 each include a Central Processing Unit (CPU) 1, a Read Only Memory (ROM) 2, and a Random Access Memory (RAM) 3. The CPU 1, the ROM 2, and the RAM 3 form what is generally referred to as a computer.
The correction model generation apparatus 400, the prediction apparatus 500, and the provisional prediction model generation apparatus 600 can each be equipped with an auxiliary memory device 4, a display device 5, an operation device 6, and an Interface (I/F) device 7. The hardware components of the correction model generation apparatus 400, the prediction apparatus 500, and the provisional prediction model generation apparatus 600 are mutually coupled through a bus 8.
The CPU 1 is an operation device configured to execute various programs installed on the auxiliary memory device 4.
The ROM 2 is a nonvolatile memory. The ROM 2 functions as a main memory device configured to store, for example, various programs and data needed for the CPU 1 to execute the various programs installed on the auxiliary memory device 4. Specifically, the ROM 2 functions as a main memory device configured to store, for example, a boot program such as a Basic Input/Output System (BIOS) or an Extensible Firmware Interface (EFI).
The RAM 3 is a volatile memory such as a Dynamic Random Access Memory (DRAM) or a Static Random Access Memory (SRAM). The RAM 3 functions as a main memory device configured to provide a work area in which various programs installed on the auxiliary memory device 4 are deployed when executed by the CPU 1.
The auxiliary memory device 4 is an auxiliary memory device configured to store various programs and information used when the various programs are executed.
The display device 5 is a display device configured to display, for example, internal statuses of the correction model generation apparatus 400, the prediction apparatus 500, and the provisional prediction model generation apparatus 600.
The operation device 6 is an input device via which administrators of the correction model generation apparatus 400, the prediction apparatus 500, and the provisional prediction model generation apparatus 600 input various instructions into the correction model generation apparatus 400, the prediction apparatus 500, and the provisional prediction model generation apparatus 600.
The I/F device 7 is a communication device configured to connect to various sensors and networks to communicate with other terminals.
<Functional Blocks>
The functional blocks of the correction model generation apparatus 400 will be described with reference to
The predicted value acquiring unit 402 is configured to acquire operational data of the device that is device type A. The predicted value acquiring unit 402 is also configured to input the operational data of the device that is device type A into the provisional prediction model (i.e., the model for the device type B) 10 to output a predicted value for an operational state.
The actually measured value acquiring unit 403 is configured to acquire an actually measured value for an operational state of the device that is device type A (i.e., an operational state calculated from the operational data of the device that is device type A). The actually measured value acquiring unit 403 can calculate an operational state from the operational data of the device that is device type A.
The training unit (correction model generation unit (which is an example of a correction unit)) 401 is configured to generate the correction model 20. Specifically, the training unit (correction model generation unit) 401 is configured to generate the correction model 20 by performing machine learning by associating the predicted value for the operational state of the device that is device type A acquired by the predicted value acquiring unit 402 with the actually measured value for the operational state of the device that is device type A acquired by the actually measured value acquiring unit 403.
The operational data acquiring unit 502 is configured to acquire operational data (specifically, operational data of the device that is device type A).
The prediction unit 501 is configured to acquire a predicted value for an operational state of the device that is device type A by inputting the operational data of the device that is device type A acquired by the operational data acquiring unit 502 into the provisional prediction model (i.e., the model for the device type B) 10 to output an operational state. The prediction unit 501 is also configured to acquire a corrected predicted value for the operational state of the device that is device type A by inputting the predicted value for the operational state of the device that is device type A into the correction model to output the corrected predicted value for the operational state.
The output unit 503 is configured to output the corrected predicted value for the operational state of the device that is device type A predicted by the prediction unit 501. Subsequently, the corrected predicted value for the operational state of the device that is device type A may be used in order to sense leakage of a refrigerant from the device that is device type A, sense a failure of the device that is device type A, or control the device that is device type A.
Here, examples of the operational data of the device will be described.
For example, the operational data of the device may include at least one selected from the following.
For example, the operational data of the device may include at least one selected from the following in addition to the operational data described above (Example 1) or instead of the operational data described above (Example 1).
For example, operational data for deducing a predicted value for an index value for a refrigerant amount during a normal operation may include either or both of the following in addition to the operational data described above (Example 1 and Example 2) or instead of the operational data described above (Example 1 and Example 2).
Here, examples of the operational state of the device will be described.
For example, the operational state may include at least one selected from the following.
For example, the value based on the supercooling degree at the outlet of the outdoor heat exchanger is a value calculated using the supercooling degree at the outlet of the outdoor heat exchanger. For example, a value calculated using the supercooling degree at the outlet of the outdoor heat exchanger is as descried below.
For example, the value based on the supercooling degree at the outlet of the outdoor heat exchanger is a value defined from physical properties of a refrigerant and refrigeration cycle diagrams (T-S and P-h diagrams).
For example, the operational state may include at least one selected from the following in addition to the index value for the refrigerant amount described above (Example 1) or instead of the supercooling degree at the outlet of the outdoor heat exchanger in the index value for the refrigerant amount described above (Example 1).
In a case of an air-warming operation, the operational state may include at least one selected from the following instead of the operational states described above (Example 1 and Example 2).
The supercooling degree at the outlet of an indoor heat exchanger is any one selected from: at least one of the supercooling degrees of the plurality of indoor heat exchangers 301; the average of the supercooling degrees of the plurality of indoor heat exchangers 301; and the supercooling degree at the indoor junction or the outdoor junction of the plurality of indoor heat exchangers 301.
In a case of a simultaneous air-cooling and warming operation, the operational state include the following in addition to the operational states described above (either or both of Example 1 and Example 2).
Examples of the device will be described below.
For example, the device used for generating the correction model 20 is the device that is the same as and of the same device type as that of the device for which the prediction apparatus 500 performs prediction.
For example, the device used for generating the correction model 20 is one or a plurality of devices different from and of the same device type as that of the device for which the prediction apparatus 500 performs prediction.
For example, the device used for generating the correction model 20 include the device that is the same as and of the same device type as that of, and one or a plurality of devices different from and of the same device type as that of, the device for which the prediction apparatus 500 performs prediction.
For example, the device type of the device is a new device type of a device, which is of a device type different from that of the device. That is, the device type A is a new device type of the device type B.
For example, the device has a function similar to that of a device, which is of a device type different from that of the device. That is, the device type A and the device type B have a similar function.
<Method>
A provisional prediction model generation process will be described below with reference to
In the step 11 (S11), the provisional prediction model generation apparatus 600 acquires training data (operational data and an operational state of a device that is device type B). When the device type B is an old device type of the device type A, and a prediction model for the device type B has already been generated, this process may be omitted by using this prediction model as a provisional model.
In the step 12 (S12), the provisional prediction model generation apparatus 600 generates a provisional prediction model (i.e., a model for the device type B) 10 by performing machine learning by using the training data acquired in S11 (specifically, by performing machine learning by associating the operational data and the operational state of the device that is device type B with each other).
In the step 21 (S21), the predicted value acquiring unit 402 acquires operational data of the device that is device type A.
In the step 22 (S22), the predicted value acquiring unit 402 inputs the operational data of the device that is device type A acquired in S21 into the provisional prediction model (i.e., the model for the device type B) 10, to output a predicted value for an operational state.
In the step 23 (S23), the actually measured value acquiring unit 403 acquires an actually measured value for the operational state of the device that is device type A (i.e., an operational state calculated from the operational data of the device that is device type A).
S23 may be performed first and S21 and S22 may be performed after S23, or S21 and S22 may be performed simultaneously with S23.
In the step 24 (S24), the training unit (correction model generation unit) 401 generates a correction model 20. Specifically, the training unit (correction model generation unit) 401 generates the correction model 20 by performing machine learning by associating the predicted value for the operational state of the device that is device type A acquired in S22 with the actually measured value for the operational state of the device that is device type A acquired in S23.
In the step 31 (S31), the operational data acquiring unit 502 acquires operational data (specifically, operational data of the device that is device type A).
In the step 32 (S32), the prediction unit 501 acquires a predicted value for an operational state of the device that is device type A by inputting the operational data of the device that is device type A acquired in S31 into the provisional prediction model (i.e., the model for the device type B) 10 to output an operational state.
In the step 33 (S33), the prediction unit 501 acquires a corrected predicted value for the operational state of the device that is device type A by inputting the predicted value for the operational state of the device that is device type A acquired in S32 into the correction model 20 to output the corrected predicted value for the operational state.
In the step 34 (S34), the output unit 503 outputs the corrected predicted value for the operational state of the device that is device type A of S33. Subsequently, the corrected predicted value for the operational state of the device that is device type A may be used to sense leakage of a refrigerant from the device that is device type A, to sense a failure of the device that is device type A, or to control the device that is device type A.
<<Updating from “Provisional Prediction Model+Correction Model” to “Prediction Model”>>
<<Updating of Correction Model>>
<Correction Model Generation Example>
The prediction apparatus 500 can predict a difference between a predicted value for an operational state of the device and an actually measured value for the operational state of the device. The details will be described with reference to
<Generation of Provisional Prediction Model>
First, the provisional prediction model generation apparatus 600 generates a provisional prediction model 10. Specifically, the provisional prediction model generation apparatus 600 generates a provisional prediction model (i.e., a model for a device B) 10 by performing machine learning by using training data (specifically, by performing machine learning by associating operational data and an operational state of the device B with each other).
<Generation of Correction Model>
Next, the correction model generation apparatus (an example of the correction apparatus) 400 generates a correction model 20. The correction model is a model configured to predict a difference between a predicted value for an operational state obtained by the provisional prediction model 10 and an actually measured value for the operational state.
Specifically, the correction model generation apparatus 400 acquires a predicted value for an operational state of the device A by inputting operational data of the device A into the provisional prediction model (i.e., the model for the device B) 10. The correction model generation apparatus 400 acquires an actually measured value for an operational state of the device A (specifically, an operational state calculated from the operational data of the device A). Moreover, the correction model generation apparatus 400 acquires operational data of the device A. Then, the correction model generation apparatus 400 generates a correction model 20 by performing machine learning by associating the predicted value for the operational state of the device A, the actually measured value for the operational state of the device A, and the operational data of the device A with one another.
<Operation Using Provisional Prediction Model and Correction Model in Combination>
Subsequently, the device A starts to be operated. The prediction apparatus 500 predicts an operational state from operational data of the device A.
Specifically, the prediction apparatus 500 acquires a predicted value for an operational state of the device A by inputting operational data of the device A into the provisional prediction model (i.e., the model for the device B) 10. Moreover, the prediction apparatus 500 acquires a predicted value for a difference between the predicted value and an actually measured value for the operational state of the device A by inputting the operational data of the device A into the correction model 20. Then, the prediction apparatus 500 acquires a corrected predicted value for the operational state based on the predicted value for the operational state of the device A, and the predicted value for the difference between the predicted value and the actually measured value for the operational state of the device A.
Embodiment 2 and Embodiment 3 will be described below. Description of any contents that are the same as those in Embodiment 1 will be omitted.
A correction apparatus 410 can correct a threshold for device abnormality determination by using a predicted value for an operational state of a device predicted using a provisional prediction model. The details will be described with reference to
<Generation of Provisional Prediction Model and Calculation of Threshold>
First, the provisional prediction model generation apparatus 600 generates a provisional prediction model 10. Specifically, the provisional prediction model generation apparatus 600 generates a provisional prediction model (i.e., a model for a device B) 10 by performing machine learning by using training data (specifically, by performing machine learning by associating operational data and an operational state of the device B with each other).
Moreover, a threshold for the device B (referred to as s B) is calculated from: a predicted value for an operational state of the device B (specifically, the operational state is output by inputting operational data of the device B into the provisional prediction model 10); and an actually measured value for the operational state of the device B (specifically, the actually measured value is an operational state calculated from the operational data of the device B). For example, where an average and a standard deviation of a Δ operational state quantity of the device B (=an actually measured value for an operational state quantity of the device B−a predicted value for the operational state quantity of the device B predicted by the provisional prediction model 10) are denoted by μ_b and σ_b, respectively, the threshold (ε_B) for the device B can be defined as “μ_b−3×σ_b”.
<Correction of Abnormality Determination Threshold>
Next, the correction apparatus 410 corrects the threshold. First, a predicted value for an operational state of the device A is acquired by inputting operational data of the device A into the provisional prediction model (i.e., the model for the device B) 10. Where an average and a standard deviation of a Δ operational state quantity of the device A (=an actually measured value for an operational state quantity of the device A−a predicted value for the operational state quantity of the device A predicted by the provisional prediction model 10) are denoted by μ_a and σ_a, respectively, a threshold (ε_A) for the device A can be defined as “μ_a−3×σ_a”. In this way, the threshold is corrected from ε_B to ε_A.
<Operation Using Provisional Prediction Model and Corrected Threshold in Combination>
Subsequently, the device A starts to be operated. An abnormality determination apparatus 510 determines abnormality from operational data of the device A. Specifically, the abnormality determination apparatus 510 determines an abnormality (e.g., leakage of a refrigerant from the device A or a failure of the device A) by comparing the Δ operational state quantity of the device A (=an actually measured value for an operational state quantity of the device A−a predicted value for the operational state quantity of the device A predicted by the provisional prediction model 10) with the threshold (ε_A) for the device A.
The correction apparatus can correct a control gain involved in controlling a device by using a predicted value for an operational state of the device predicted using a provisional prediction model. The details will be described with reference to
<Generation of Provisional Prediction Model and Calculation of Control Gain>
First, the provisional prediction model generation apparatus 600 generates a provisional prediction model 10. Specifically, the provisional prediction model generation apparatus 600 generates a provisional prediction model (i.e., a model for a device B) 10 by performing machine learning by using training data (specifically, by performing machine learning by associating operational data and an operational state of the device B with each other).
Moreover, a control gain of the device B (an output from the device B (an actually measured value for an operational state)/an input into the device B (an actually measured value for operational data)) is calculated. The control gain of the device B is referred to as “K_B”.
<Addition of Correction Control Gain>
Next, the correction apparatus 410 calculates a correction control gain to be added. A correction control gain (a correction coefficient for an output from the provisional prediction model (i.e., the model for the device B) 10) is referred to as “K_c”. The correction control gain (K_c) is “an output (an actually measured value for an operational state) from the device A/an output (a predicted value for the operational state) from the provisional prediction model (i.e., the model for the device B) 10” with respect to the same input. Hence, a predicted value for the output from the device A can be calculated according to K_c×the output from the provisional prediction model (i.e., the model for the device B) 10.
<Operation Using Provisional Prediction Model and Correction Control Gain in Combination (Example of Application to Internal Model Control (IMC)>
Subsequently, the device A starts to be operated. A device control apparatus 520 controls the device A by using the provisional prediction model (i.e., the model for the device B) 10 and the correction control gain (K_c), which is the correction coefficient for an output from the provisional prediction model (i.e., the model for the device B) 10. Specifically, the device control apparatus 520 performs control such that an output from the device A (an actually measured value for an operational state) becomes closer to a target value. In the present embodiment, a gain for correcting an output from the provisional prediction model (i.e., the model for the device B) 10 is added. However, a gain for correcting the control gain K_B of the device B may be added.
Embodiments have been described above. It will be understood that various modifications are applicable to the embodiments and particulars without departing from the spirit and scope of the claims.
Number | Date | Country | Kind |
---|---|---|---|
2021-044569 | Mar 2021 | JP | national |
This application is a continuation application of International Application No. PCT/JP2022/012808, filed on Mar. 18, 2022, and designating the U.S., which is based upon and claims priority to Japanese Patent Application No. 2021-044569, filed on Mar. 18, 2021, the entire contents of which are incorporated herein by reference.
Number | Date | Country | |
---|---|---|---|
Parent | PCT/JP2022/012808 | Mar 2022 | US |
Child | 18467090 | US |