Parameter search device and parameter search method

Abstract

A parameter search device (1) is provided with: an operation result collection unit (11) that collects operation results including parameters indicating operation conditions of a mechanical device (2); an evaluation value acquisition unit (12) that acquires an evaluation value of the parameter obtained using the operation result; and a parameter search unit (13) that searches for a target parameter from among parameters indicating the operating conditions of the mechanical device (2). The parameter search unit (13) predicts the relationship between the parameter and the evaluation value using a machine learning model, and determines whether or not to continue the search of the parameter on the basis of the result of comparison between a search end index and a threshold value, using the number of search candidates for which the evaluation value exceeds the optimal search as the search end index.

Description

Technical Field

[0001] The present disclosure relates to a parameter search device and a parameter search method. Background Art

[0002] A technology for searching for parameters that define operating conditions of a mechanical device according to an environment in which the mechanical device is installed or each operating task of the mechanical device is proposed. For example, Patent Document 1 describes the following air conditioner linkage control system: an air conditioner is provided, and in a data center that processes data through an information processing device, the setting conditions of the air conditioner are controlled in linkage with the load distribution of a CPU (Central Processing Unit) of the information processing device.

[0003] The air conditioner linkage control system described in Patent Document 1 searches for the optimal combination state of the combination of the setting conditions of the air conditioner and the load distribution of the CPU of the information processing device that minimizes the estimated power consumption estimated by the power consumption learning model under the condition that the estimated temperature and humidity estimated by the temperature and humidity learning model fall within a predetermined specified range.

[0004] Prior art literature

[0005] Patent Literature

[0006] Patent Document 1: Japanese Patent Application Publication No. 2012-149839 Summary of the invention

[0007] Problem that the invention aims to solve

[0008] The system described in Patent Document 1 searches for parameters representing the operating conditions of the air conditioner and the information processing device as mechanical devices under the condition that the estimated temperature and humidity fall within a predetermined range. Therefore, the parameter search is performed according to a fixed search number or search time determined by the above-mentioned predetermined range, so there is a problem that the parameter cannot be searched efficiently. For example, there may be a case where the target parameter cannot be fully searched according to the search number or search time, or even if the target parameter is searched, unnecessary search continues.

[0009] The present disclosure is to solve the above-mentioned problem, and an object of the present disclosure is to obtain a parameter search device and a parameter search method capable of efficiently searching for parameters representing target operating conditions of a mechanical device.

[0010] Means used to solve problems

[0011] The parameter search device disclosed in the present invention comprises: an operation result collecting unit, which collects operation results including parameters representing the operation conditions of a mechanical device; an evaluation value acquiring unit, which acquires the evaluation values ​​of the parameters obtained using the operation results; and a parameter searching unit, which searches for target parameters from the parameters representing the operation conditions of the mechanical device, the parameter searching unit using a machine learning model to predict the relationship between the parameters in the search candidates and the evaluation values, the machine learning model outputting a predicted value representing the relationship between the parameters in the search candidates and the evaluation values ​​when a plurality of search candidates regarding the parameters and the evaluation values ​​are input, the predicted result of the machine learning model is compared with the best search with the largest evaluation value among the searched parameters, the number of search candidates with evaluation values ​​exceeding the best search is used as a search end indicator, and based on the comparison result of the search end indicator with a threshold value, it is determined whether to continue the parameter search.

[0012] Effects of the Invention

[0013] According to the present disclosure, the result of predicting the relationship between the parameter as a search candidate and the evaluation value is compared with the best search with the largest evaluation value among the searched parameters, and based on the comparison result of the number of search candidates whose evaluation values ​​exceed the best search, i.e., the search end index, and the threshold, it is determined whether to continue the parameter search. Thus, the parameter search device of the present disclosure can efficiently search for parameters representing the target operating conditions of the mechanical device without having to determine a fixed number of searches or search time in advance. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a block diagram showing the structure of the parameter search device according to the first embodiment.

[0015] Figure 2 It is a diagram showing the relationship between parameters and evaluation values ​​in the initial stage of parameter search.

[0016] Figure 3 It is a diagram showing the relationship between the parameters and the evaluation values ​​in the late stage of the parameter search.

[0017] Figure 4 : is a flowchart showing the parameter search method of embodiment 1.

[0018] Figure 5A and Figure 5B This is a block diagram showing a hardware configuration for realizing the functions of the parameter search device according to the first embodiment.

[0019] Figure 6 This is a block diagram showing the structure of a specific example of the parameter search device according to the first embodiment.

[0020] Figure 7 This is a flowchart showing a specific example of the parameter search method according to the first embodiment.

[0021] Figure 8 This is a block diagram showing the structure of a parameter search device according to the second embodiment.

[0022] Fig. 9 This is a block diagram showing the structure of a parameter search device according to a third embodiment.

[0023] Fig.10 This is a flowchart showing the parameter search method of the third embodiment.

[0024] Fig.11 This is a block diagram showing the structure of a parameter search device according to a fourth embodiment.

[0025] Fig.12 This is a block diagram showing the structure of a parameter search device according to a fifth embodiment.

[0026] Fig.13 This is a conceptual diagram showing a refrigeration cycle of the air-conditioning cooling and heating device in the fifth embodiment.

[0027] Fig.14 This is a flowchart showing the parameter search method of embodiment 5.

[0028] Fig.15 This is a block diagram showing the structure of an air conditioning and cooling device including a parameter search device according to a fifth embodiment.

[0029] Fig.16 This is a block diagram showing the structure of a parameter search device according to a sixth embodiment.

[0030] Fig.17 This is a conceptual diagram showing a refrigeration cycle of an air-conditioning and cooling device in a sixth embodiment. DETAILED DESCRIPTION

[0031] Implementation method 1.

[0032] Figure 1 is a block diagram showing the structure of the parameter search device 1 according to the first embodiment. Figure 1 In the embodiment, the parameter search device 1 is a device for searching for target operating conditions of the mechanical device 2, specifically, searching for parameters indicating operating conditions of the mechanical device 2. The value of the parameter found in the search result is set in the mechanical device 2. The mechanical device 2 is, for example, an air conditioning and heating device including a fan, an electronic expansion valve, a compressor, etc., an industrial device such as a machine tool, or a household electronic product.

[0033] The operating conditions refer to information indicating the content of the actions performed by the various structural elements of the control object possessed by the mechanical device 2. For example, when the mechanical device 2 is an air conditioning and heating device, the operating conditions are the values ​​of the operating parameters of the electronic expansion valve, solenoid valve, and compressor possessed by the air conditioning and heating device, the air volume of the fan, the speed of the fan, the angle of the blades that determine the blowing direction, and the flow rate of the water to be controlled for heating or cooling.

[0034] In the prior art of controlling the operation of a mechanical device, a parameter representing the operating condition of the mechanical device is searched under the condition that the estimated peripheral state of the mechanical device falls within a predetermined range. In this case, the parameter search is performed according to a fixed search number or search time determined by the predetermined range. Therefore, there may be a case where the target parameter cannot be fully searched depending on the search number or search time, or even if the target parameter is searched, unnecessary search is continued.

[0035] When the optimal value of the evaluation value of the parameter representing the operating condition of the mechanical device is known, an index called simple regret can be used as shown in Reference 1. The termination of the parameter search can be determined based on the simple regret, thereby shortening the parameter search time.

[0036] However, if the optimal value of the evaluation value of the parameter is not known, this method cannot be used for parameter search.

[0037] (Reference 1) Automatic Termination for Hyperparameter Optimization, ICLR2022

[0038] In contrast, the parameter search device 1 uses a machine learning model to predict the relationship between parameters and evaluation values ​​in multiple search candidates. Then, the parameter search device 1 compares the prediction result of the machine learning model with the best search with the largest evaluation value among the searched parameters, and uses the number of search candidates with evaluation values ​​exceeding the best search as the search end indicator. The parameter search device 1 determines whether to continue the parameter search based on the comparison result of the search end indicator and the threshold.

[0039] For example, the parameter search device 1 uses Bayesian optimization to efficiently search for parameters representing the optimal operating conditions of the mechanical device 2, and uses the predicted relationship between the parameters and the evaluation value based on the Gaussian process regression model to determine the search end index. By using the determined search end index, the parameter search is ended at an appropriate time. As a result, in the parameter search device 1, even if a fixed number of searches or search time is not set in advance, it is possible to efficiently search for parameters representing the target operating conditions of the mechanical device 2.

[0040] like Figure 1 As shown, the parameter search device 1 includes an operation result collecting unit 11, an evaluation value obtaining unit 12, and a parameter search unit 13. The operation result collecting unit 11 collects the operation result of the mechanical device 2. The operation result of the mechanical device 2 is information including parameters indicating the operation conditions of the mechanical device 2. The operation result is, for example, the value of the parameter set as the operation condition of the mechanical device 2, information indicating the structure of the mechanical device 2, information inherent to the mechanical device 2 such as the operation mode, and information including log data of past operation.

[0041] The operation result collection unit 11 may also obtain the operation result of the mechanical device 2 from the mechanical device 2 via the encoder. For example, the operation result collection unit 11 may directly collect detection values ​​detected by a sensor such as a temperature sensor, a pressure sensor, an acceleration sensor, a gyro sensor, or a humidity sensor provided in the mechanical device 2, or may collect calculated values ​​using these detection values.

[0042] The evaluation value acquisition unit 12 acquires the evaluation value obtained using the operation result for the parameter representing the operation condition. The evaluation value is an index value indicating how close the parameter corresponding to the evaluation value is to the parameter representing the target operation condition of the mechanical device 2. For example, the evaluation value and the parameter may be stored in advance in a storage device from which the parameter search device 1 can read the stored content, and the evaluation value acquisition unit 12 reads and acquires the evaluation value corresponding to the parameter included in the operation result from the storage device.

[0043] In addition, the evaluation value of the parameter may be a value that is maximized or minimized in such a way that the difference between the measured parameter value and its target value is minimized. For example, when there is a target value for the motion speed of the structural element of the mechanical device 2, the evaluation value acquisition unit 12 acquires the evaluation value represented by the following formula (1), and the parameter search device 1 searches for the parameter with the minimum evaluation value as the target parameter.

[0044] Evaluation value = (actually measured motion speed - target motion speed) 2 ···(1)

[0045] When there are a plurality of parameters for one operating condition of the mechanical device 2 , the evaluation value acquisition unit 12 may acquire evaluation values ​​of each of these parameters, or may acquire a common evaluation value obtained by performing a weighted sum operation or the like.

[0046] In addition, the evaluation value acquisition unit 12 may also calculate the evaluation value of the parameter included in the operation result using the operation result of the mechanical device 2. For example, when a target value is set for the motion speed of a component of the mechanical device 2 and the actual motion speed is obtained as the operation result, the evaluation value acquisition unit 12 uses both to calculate the evaluation value shown in the above formula (1).

[0047] In this way, “obtaining” the evaluation value of the parameter in the parameter search device 1 also includes obtaining the evaluation value by “calculation” using the operation result of the mechanical device 2 .

[0048] The parameter search unit 13 searches for a target parameter from the parameters representing the operating conditions of the mechanical device 2 collected by the operating result collection unit 11. Specifically, the parameter search unit 13 predicts the relationship between the parameters and the evaluation values ​​in the search candidates using the following machine learning model, which outputs the relationship between the parameters and the evaluation values ​​in the search candidates when a plurality of search candidates regarding the parameters and the evaluation values ​​are input.

[0049] For machine learning models, one or more models of various models such as linear regression, generalized linear model, Gaussian process regression, hierarchical Bayesian model, neural network, neural process, random forest or gradient boosting tree are used.

[0050] In addition, when the number of dimensions input to the machine learning model, that is, the number of input parameters, is large, the parameters after the number of dimensions has been reduced by methods such as principal component analysis, singular value decomposition, tensor decomposition, or Auto Encoder, which are methods of reducing the number of dimensions, can also be input into the machine learning model.

[0051] For example, the parameter search unit 13 uses the search candidates of the parameters generated based on the operation results of the mechanical device 2 as learning data, and causes the machine learning model to learn the relationship between the parameters and the evaluation values ​​included in the operation results. In this way, the parameter search unit 13 generates the machine learning model using the measured data of the operation results of the mechanical device 2 each time, so that the machine learning model corresponding to the current status of the mechanical device 2 can be obtained.

[0052] Alternatively, the machine learning model and the parameters may be stored in advance in a storage device from which the parameter search device 1 can read the stored contents, and the parameter search unit 13 may read the machine learning model corresponding to the parameters included in the operation result from the storage device for use.

[0053] The parameter search unit 13 compares the prediction result of the machine learning model with the best search with the largest evaluation value among the searched parameters, and sets the number of search candidates whose evaluation values ​​exceed the best search as the "search end index". The parameter search unit 13 determines whether to continue the parameter search based on the comparison result of the search end index and the threshold. The parameters of the final search result are set in the mechanical device 2. As a result, the mechanical device 2 operates according to the target operating conditions indicated by the parameters of the search result.

[0054] Figure 2 It is a diagram showing the relationship between parameters and evaluation values ​​in the initial stage of parameter search. Figure 3 is a diagram showing the relationship between parameters and evaluation values ​​in the later stage of parameter search. Figure 2 and Figure 3 In the example, the horizontal axis is a parameter representing a certain operating condition, and the vertical axis is an evaluation value of the parameter. The parameter search unit 13 uses the parameters and evaluation values ​​included in the operating results of the mechanical device 2 collected by the operating result collection unit 11 as learning data, and enables the machine learning model to learn the relationship between the parameters and evaluation values ​​included in the operating results. Next, the parameter search unit 13 generates a plurality of search candidates for the parameters and evaluation values, and predicts the relationship between the parameters and evaluation values ​​in the search candidates by inputting these search candidate data into the machine learning model.

[0055] The prediction result of the relationship between the parameter and the evaluation value in the search candidate is, for example, the predicted value of the average μ and the standard deviation σ of the search candidate data using the parameter and the evaluation value as elements. Figure 2 and Figure 3 In the figure, A1 and A2 are functions representing the average μ of the evaluation value for the parameter. B1 and B2 are functions representing +2σ about the average μ. C is the parameter and evaluation value collected by the operation result collection unit 11, which is the actual measurement data measured in the mechanical device 2. The average μ shown in functions A1 and A2 is predicted for the actual measurement point C using the machine learning model. D1 and D2 are functions representing -2σ about the average A1 and A2.

[0056] The parameter search unit 13 uses the measured value of the parameter with the largest evaluation value among the parameters included in the measured operation results as the optimal search E1. Figure 2 The number of search candidates of the best search E1 shown by the dotted line is counted, and the counted number is used as the search end indicator. In the initial stage of parameter search, as shown in FIG. Figure 2 As shown, the number of search candidates exceeding the optimal search E1 is large, so the value of the search end index becomes greater than the threshold. When the search end index is greater than the threshold, the parameter search unit 13 determines to continue the parameter search.

[0057] When the parameter search is continued, the parameter search unit 13 determines, for example, a search candidate whose predicted parameter and evaluation value relationship satisfies the selection condition as the next search object and sets it to the mechanical device 2. For example, a search candidate whose parameter and evaluation value relationship is closest to the target is determined as the next search object. The parameters of the determined search object are set to the mechanical device 2. Thus, the parameter search unit 13 repeatedly performs the above-mentioned series of parameter searches by operating the mechanical device 2 under the operating conditions indicated by the parameters of the search object and collecting the operating results by the operating result collection unit 11. Repeating the parameter search, transfer to Figure 3 The parameter search is shown in the later stages.

[0058] The parameter search unit 13 uses the measured value of the parameter with the largest evaluation value among the parameters included in the measured operation results as the optimal search E2. Figure 3 The number of search candidates of the best search E2 shown by the dotted line is counted, and the counted number is used as the search end indicator. Figure 3 As shown, the number of search candidates exceeding the optimal search E2 decreases. For example, when the search end index is less than the threshold, the parameter search unit 13 determines the end of the parameter search. When the threshold is "0", the parameter search ends when the search end index becomes 0, that is, when the number of search candidates exceeding the optimal search becomes 0.

[0059] In this way, the parameter search device 1 can efficiently perform the parameter search by combining the parameter search and the search end determination.

[0060] Furthermore, the parameter search can be terminated at an appropriate timing based on the search end indicator, and thus the parameter search time can be shortened.

[0061] Figure 4 1 is a flowchart showing a parameter search method according to Embodiment 1, which is a parameter search method performed by the parameter search device 1 .

[0062] The operation result collecting unit 11 collects the operation results including the parameters indicating the operation conditions of the mechanical device 2 (step ST1 ). The evaluation value acquiring unit 12 acquires the evaluation values ​​of the parameters obtained using the operation results (step ST2 ).

[0063] Next, the parameter search unit 13 searches for target parameters from among the parameters indicating the operating conditions of the mechanical device 2 (step ST3 ).

[0064] In step ST3, the parameter search unit 13 uses a machine learning model to predict the relationship between the parameters and evaluation values ​​in the search candidates, compares the prediction results of the machine learning model with the best search with the largest evaluation value among the searched parameters, takes the number of search candidates with evaluation values ​​exceeding the best search as a search end indicator, and determines whether to continue the parameter search based on the comparison result between the search end indicator and the threshold.

[0065] By executing the above-described parameter search method, even without determining a fixed number of searches or a fixed search time, it is possible to efficiently search for parameters representing target operating conditions of the mechanical device 2 .

[0066] Next, the hardware configuration for realizing the functions of the parameter search device 1 will be described.

[0067] The functions of the operation result collecting unit 11, the evaluation value acquiring unit 12, and the parameter searching unit 13 of the parameter searching device 1 are realized by the processing circuit. Figure 4 The processing circuit for the processes of step ST1 to step ST3 shown in the figure may be dedicated hardware, or may be a CPU that executes a program stored in a memory.

[0068] Figure 5A 2 is a block diagram showing a hardware configuration for realizing the functions of the parameter search device 1 . Figure 5B 1 is a block diagram showing a hardware structure for executing software that realizes the functions of the parameter search device 1. Figure 5A and Figure 5B In FIG. 1 , the input interface 100 is an interface for the parameter search device 1 to relay information indicating the operation result from the mechanical device 2 . The output interface 101 is an interface for relaying parameters of the search result output from the parameter search device 1 to the mechanical device 2 .

[0069] The processing circuit is Figure 5A In the case of the dedicated hardware processing circuit 102 shown, the processing circuit 102 corresponds to, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination thereof. The functions of the operation result collecting unit 11, the evaluation value acquiring unit 12, and the parameter searching unit 13 of the parameter searching device 1 may be realized by different processing circuits, or these functions may be realized by one processing circuit.

[0070] The processing circuit is Figure 5BIn the case of the processor 103 shown, the functions of the operation result collecting unit 11, the evaluation value acquiring unit 12 and the parameter searching unit 13 of the parameter searching device 1 are realized by software, firmware or a combination of software and firmware. The software or firmware is described in the form of a program and stored in the memory 104.

[0071] The processor 103 realizes the functions of the operation result collecting unit 11, the evaluation value obtaining unit 12, and the parameter searching unit 13 of the parameter searching device 1 by reading and executing the program stored in the memory 104. For example, the parameter searching device 1 has a memory 104 for storing a program, and when the program is executed by the processor 103, the result is executed. Figure 4 The programs cause the computer to execute the process or method of the processing performed by the operation result collection unit 11, the evaluation value acquisition unit 12, and the parameter search unit 13. The memory 104 may also be a computer-readable storage medium storing a program for causing the computer to function as the operation result collection unit 11, the evaluation value acquisition unit 12, and the parameter search unit 13.

[0072] The memory 104 corresponds to, for example, RAM (Random Access Memory), ROM (Read Only Memory), flash memory, EPROM (Erasable Programmable ROM), EEPROM (Electrically-EPROM) (registered trademark), and other non-volatile or volatile semiconductor memories, magnetic disks, floppy disks, optical disks, high-density disks, mini disks, DVDs, and the like.

[0073] Alternatively, the function of the operation result collection unit 11, the evaluation value acquisition unit 12, and the parameter search unit 13 of the parameter search device 1 may be partially realized by dedicated hardware, and the other part may be realized by software or firmware. For example, the operation result collection unit 11 and the evaluation value acquisition unit 12 may be realized by the processing circuit 102 as dedicated hardware, and the parameter search unit 13 may be realized by the processor 103 reading and executing the program stored in the memory 104. In this way, the processing circuit may realize the above-mentioned functions by hardware, software, firmware, or a combination thereof.

[0074] Hereinafter, a specific example of the parameter search device 1 will be described.

[0075] Figure 6 1 is a block diagram showing a specific example of the structure of the parameter search device 1. Figure 6In the embodiment, the parameter search device 1 includes an operation result collection unit 11, an evaluation value acquisition unit 12, and a parameter search unit 13. The operation result collection unit 11 includes an operation result acquisition unit 111 and an operation result storage unit 112. The evaluation value acquisition unit 12 includes an evaluation value calculation unit 121 and an evaluation value storage unit 122. The parameter search unit 13 includes a machine learning unit 131, a search end determination unit 132, and a parameter determination unit 133.

[0076] The operation result acquisition unit 111 acquires the operation result of the mechanical device 2. For example, the operation result acquisition unit 111 acquires the detection values ​​detected by the sensor group (temperature sensor, pressure sensor, acceleration sensor, gyro sensor or humidity sensor) provided in the mechanical device 2. These detection values ​​are the values ​​of the parameters that define the operation conditions for operating the mechanical device 2. In addition, the operation result acquisition unit 111 may also acquire the physical quantity of the prescribed operation condition calculated using these detection values.

[0077] The operation result storage unit 112 stores the operation result acquired by the operation result acquisition unit 111. The operation result storage unit 112 is a storage device provided in a computer functioning as the parameter search device 1, and includes a storage device such as a HDD (Hard Disk Drive) or an SSD (Solid State Drive), or Figure 5B Memory 104, etc.

[0078] The evaluation value calculation unit 121 calculates the evaluation value of the parameter indicating the operating condition using the data indicating the operating result of the mechanical device 2. For example, when there is a target value for the motion speed of the structural element of the mechanical device 2, the evaluation value calculation unit 121 calculates the evaluation value according to the above formula (1) using the motion speed data included in the operating result read from the operating result storage unit 112. The evaluation value calculated by the evaluation value calculation unit 121 is stored in the evaluation value storage unit 122 in correspondence with the parameter.

[0079] The evaluation value storage unit 122 stores the evaluation values ​​of the parameters. The evaluation value storage unit 122 is a storage device provided in a computer functioning as the parameter search device 1, and includes a storage device such as a HDD or SSD, or Figure 5B Memory 104, etc.

[0080] The machine learning unit 131 generates a machine learning model for predicting the relationship between the parameters and the evaluation values ​​in the search candidates based on the operation results of the mechanical device 2, and uses the generated machine learning model to predict the relationship between the parameters and the evaluation values. Figure 6 As shown, the machine learning unit 131 includes a learning prediction unit 1311 and a prediction result storage unit 1312 .

[0081] The learning prediction unit 1311 generates search candidates for parameters and evaluation values ​​based on the operation results of the mechanical device 2, and uses the search candidates as learning data to make the machine learning model learn the relationship between the parameters and evaluation values ​​included in the operation results. In addition, the learning prediction unit 1311 uses the machine learning model to predict the relationship between the parameters and evaluation values ​​in the search candidates.

[0082] The prediction result storage unit 1312 stores the prediction results of the machine learning model. The prediction result storage unit 1312 is a storage device provided in the computer functioning as the parameter search device 1, and includes a storage device such as a HDD or SSD, or Figure 5B Memory 104, etc.

[0083] The search end determination unit 132 determines the end of the parameter search based on the prediction result of the machine learning model. For example, the search end determination unit 132 sets the measured value of the parameter with the largest evaluation value in the operation result as the best search, and sets the number of search candidates that exceed the best search as the search end index. For example, when the search end index is above a threshold, the search end determination unit 132 determines to continue the parameter search, and when the search end index is less than the threshold, it determines to end the parameter search. When it is determined to continue the parameter search, the search end determination unit 132 reads the prediction result of this time from the prediction result storage unit 1312 and outputs it to the parameter determination unit 133.

[0084] When the parameter determination unit 133 determines that the parameter search should continue based on the prediction result of the machine learning model, it determines the next parameter to be searched (search object). Figure 6 As shown, the parameter determination unit 133 includes a search parameter calculation unit 1331 and a parameter instruction unit 1332 .

[0085] The search parameter calculation unit 1331 calculates the next parameter to be searched. For example, the search parameter calculation unit 1331 calculates the search candidate whose relationship between the parameter and the evaluation value is closest to the target among the prediction results of the relationship between the parameter and the evaluation value this time as the next search object. In the case where Bayesian optimization is used for parameter search, the search parameter calculation unit 1331 sets the search candidate whose value of the acquisition function considering the average μ and standard deviation σ of the search candidates is the largest as the next search object.

[0086] The parameter instruction unit 1332 indicates the parameters calculated by the search parameter calculation unit 1331 to the mechanical device 2. For example, the parameter instruction unit 1332 communicates with the mechanical device 2 via a wireless or wired communication device ( Figure 6 The search target parameters are sent to the mechanical device 2 (not shown). The mechanical device 2 operates according to the operating conditions indicated by the parameters received from the parameter instruction unit 1332.

[0087] Furthermore, although the operation result storage unit 112 , the evaluation value storage unit 122 , and the prediction result storage unit 1312 are shown as separate storage units, these storage units may be provided in a storage area of ​​one storage device.

[0088] Furthermore, the operation result storage unit 112, the evaluation value storage unit 122, and the prediction result storage unit 1312 are storage units provided in the parameter search device 1, but they may be storage units provided in an external storage device provided separately from the parameter search device 1. In addition, the external storage device is a storage device that can read the stored content from the parameter search device 1 through wireless or wired communication.

[0089] Figure 7 is a flowchart showing a specific example of the parameter search method of embodiment 1, showing Figure 6 The operations of the parameter search device 1 and the mechanical device 2 are shown.

[0090] First, the search parameter calculation unit 1331 determines the initial value (initial point) of the parameter to be searched (step ST1A). For example, the search parameter calculation unit 1331 randomly determines the initial value of the parameter to be searched from the parameters representing various operating conditions of the mechanical device 2. In addition, when a parameter with a high evaluation value is known in advance, the parameter may be set as the initial value.

[0091] Next, the parameter instruction unit 1332 instructs the mechanical device 2 to set the initial value of the search target parameter calculated by the search parameter calculation unit 1331 (step ST2A). For example, the parameter instruction unit 1332 sends the initial value of the search target parameter to the mechanical device 2 via a wireless or wired communication device.

[0092] When the mechanical device 2 receives the initial value of the parameter from the parameter instruction unit 1332, it operates according to the operating condition indicated by the received parameter value (step ST3A). When the mechanical device 2 starts to operate, the operation result acquisition unit 111 sequentially acquires the operation results from the mechanical device 2 (step ST4A). The operation results acquired by the operation result acquisition unit 111 are stored in the operation result storage unit 112. Next, the evaluation value calculation unit 121 calculates the evaluation value of the parameter included in the operation result stored in the operation result storage unit 112. The evaluation value calculated by the evaluation value calculation unit 121 is stored in the evaluation value storage unit 122 in correspondence with the parameter.

[0093] Next, the learning prediction unit 1311 uses the parameters and evaluation values ​​stored in the evaluation value storage unit 122 as learning data to make the machine learning model learn the relationship between the parameters and the evaluation values ​​(step ST5A). For example, the evaluation value storage unit 122 stores the parameter x obtained before the search number t. t, state quantity s representing the state of the mechanical device 2 t And the evaluation value y t Data D as an element t ={(x1,s1,y1),···,(x t ,s t ,y t )}.

[0094] The learning prediction unit 1311 learns the data D read from the evaluation value storage unit 122 using a machine learning model. t The parameter x shown t , state quantity s representing the state of the mechanical device 2 t And the evaluation value y t In addition, as learning data, the parameter x t and the evaluation value y t is necessary, but the state quantity s t Unrestricted.

[0095] The learning prediction unit 1311 generates search candidates (search candidate points) for parameters representing the operating conditions of the mechanical device 2 (step ST6A). For example, the learning prediction unit 1311 randomly generates a plurality of search candidate points using information representing the parameters and evaluation values ​​stored in the evaluation value storage unit 122. In addition, when the maximum value or minimum value of the evaluation value of the parameter is known, the learning prediction unit 1311 may use the known evaluation value and parameter as reference points to determine the search candidate points at equal intervals in a grid-like manner as shown in the figure. In addition, the learning prediction unit 1311 may generate a plurality of search candidate points using an experimental planning method.

[0096] Next, the learning prediction unit 1311 uses the machine learning model to predict the relationship between the parameters and the evaluation value at the search candidate point (step ST7A). For example, when Bayesian optimization is used to determine the parameters of the search object, the learning prediction unit 1311 calculates the next search object by calculating the function value called the acquisition function.

[0097] The search end determination unit 132 determines whether to end the parameter search based on the prediction results stored in the prediction result storage unit 1312 (step ST8A).

[0098] For example, the search end determination unit 132 calculates the prediction results of the Gaussian process regression model as a machine learning model, that is, the average μ, standard deviation σ, and optimal search y of the search candidate points. best The comparison will exceed the optimal search y best The number of search candidate points is used as the search end indicator S t .

[0099] Search end indicator St It is calculated using the calculation formula of each acquisition function.

[0100] For example, using the UCB search end indicator S t It is expressed by the following formula (2). In addition, the search end indicator S of EI is used t It is represented by the following equation (3). In the following equations (2) and (3), COUNT[] is a function for finding the number of search candidate points that satisfy the conditions in []. The search end determination unit 132 sets the search end indicator S t = 0, it is expected that the optimal search y is not exceeded. best The search candidate points are thus determined, and therefore the termination of the parameter search can be determined.

[0101] Although the case where the threshold value is 0 is shown, any numerical value may be set as the threshold value as long as it is a value greater than 0.

[0102] S t =COUNT[(μ+κσ-y best )≥0] (2)

[0103] S t =COUNT[(μ-y best ) / σ≥0] (3)

[0104] When the search end indicator S is determined t If the value is less than the threshold value (step ST8A; Yes), the search end determination unit 132 determines that the search is an interpretation parameter search. Figure 7 On the other hand, when it is determined that the search end indicator S t If it is equal to or greater than the threshold value (step ST8A; No), the search end determination unit 132 determines to continue the parameter search, reads the prediction result of this time from the prediction result storage unit 1312 , and outputs it to the parameter determination unit 133 .

[0105] The search parameter calculation unit 1331 calculates the search candidate point with the largest value of the acquisition function taking into account the average μ and standard deviation σ of the search candidate points as the next search object (step ST9A). The parameter instruction unit 1332 indicates the parameters of the search object calculated by the search parameter calculation unit 1331 to the mechanical device 2. The mechanical device 2 operates according to the operating conditions indicated by the parameters received from the parameter instruction unit 1332. After that, the process returns to step ST2A and the above process is repeated.

[0106] As calculation methods for the search object, there are Upper Confidence Bound (UCB), and Expected Improvement (EI). The search parameter calculation unit 1331 uses the average μ and standard deviation σ of the search candidate points to calculate the search result according to the function acq represented by the following equation (4): UCB To calculate UCB. In the following formula (4), κ is a hyperparameter. When the value of the hyperparameter κ increases, the tendency to search for parameters that cannot be searched becomes stronger.

[0107] acq UCB (x)=μ+κσ (4)

[0108] The search parameter calculation unit 1331 may also use the average μ and standard deviation σ of the search candidate points according to the function acq represented by the following equation (5): EI To calculate EI. In the following formula (5), y best is the best search at a certain point in time. Z is (μ-y best ) / σ. Φ(Z) is the cumulative density function of the standard normal distribution, is the probability density function of the standard normal distribution. The search parameter calculation unit 1331 determines the search candidate point with the largest value of these acquisition functions as the next search target.

[0109]

[0110] As described above, the parameter search device 1 of the first embodiment includes an operation result collection unit 11 for collecting operation results, an evaluation value acquisition unit 12 for acquiring an evaluation value of a parameter obtained using the operation results, and a parameter search unit 13 for searching for target parameters. The parameter search unit 13 uses a machine learning model to predict the relationship between the parameter and the evaluation value, and uses the number of search candidates whose evaluation values ​​exceed the optimal search as a search end indicator. When the search for the parameter is continued based on the comparison result between the search end indicator and the threshold, the next search object is determined from the search candidates based on the evaluation value. As a result, the parameter search device 1 can efficiently search for parameters representing the target operating conditions of the mechanical device 2 without determining a fixed number of searches or search time.

[0111] In the parameter search device 1 of the first embodiment, the parameter search unit 13 trains the machine learning model. The parameter search unit 13 generates the machine learning model using the actual measurement data of the operation result of the mechanical device 2 each time, so that the machine learning model corresponding to the current status of the mechanical device 2 can be obtained.

[0112] The parameter search method of the first embodiment comprises the following steps: an operation result collecting unit 11 collects operation results; an evaluation value obtaining unit 12 obtains the evaluation value of the parameter obtained using the operation results; and a parameter search unit 13 searches for a target parameter. The parameter search unit 13 uses a machine learning model to predict the relationship between the parameter and the evaluation value, and uses the number of search candidates whose evaluation values ​​exceed the optimal search as a search end index. When the search for the parameter is continued based on the comparison result between the search end index and the threshold, the next search object is determined from the search candidates based on the evaluation value. By executing this method, the parameters representing the target operation conditions of the mechanical device 2 can be efficiently searched without determining a fixed number of searches or search time.

[0113] Implementation method 2.

[0114] In the first embodiment, the relationship between the parameters and the evaluation values ​​representing the operating conditions of the mechanical device 2 is learned, and the parameters are searched based on these relationships. In contrast, in the second embodiment, in addition to collecting parameters and evaluation values, data related to the constraints that should be satisfied in the operation of the mechanical device 2 is also collected, and the parameters are searched based on the relationship between the parameters, evaluation values, and constraints.

[0115] Figure 8 1A is a block diagram showing the structure of a parameter search device 1A according to Embodiment 2. Figure 8 In the embodiment, the parameter search device 1A includes an operation result collection unit 11A, an evaluation value acquisition unit 12A, and a parameter search unit 13A. The operation result collection unit 11A includes an operation result acquisition unit 111A and an operation result storage unit 112A. The evaluation value acquisition unit 12A includes an evaluation value calculation unit 121A, an evaluation value storage unit 122A, and a constraint calculation unit 123. The parameter search unit 13A includes a machine learning unit 131A, a search end determination unit 132A, and a parameter determination unit 133A.

[0116] The operation result acquisition unit 111A acquires the operation result including the constraint conditions during the operation of the mechanical device 2 in addition to the parameters indicating the operation conditions of the mechanical device 2. The constraint conditions are conditions that should be satisfied during the operation of the mechanical device 2, and the constraint conditions must be satisfied during the operation of the mechanical device 2, which is different from the evaluation value of the parameter.

[0117] For example, when the constraint condition is a value that a parameter should satisfy, the operation result acquisition unit 111A may acquire, as the constraint condition, a detection value detected by a sensor group (temperature sensor, pressure sensor, acceleration sensor, gyro sensor, or humidity sensor) provided in the mechanical device 2. In addition, the operation result acquisition unit 111A may acquire, as the constraint condition, a physical quantity of a predetermined operation condition calculated using these detection values.

[0118] For example, assuming that the actual measured value of the parameter representing the operating condition of the mechanical device 2 is c measure , let the threshold that the parameter should meet be c thre In the case of , the constraint condition c can be calculated using the following formula (6). The constraint condition c expressed by the following formula (6) is the actual measurement value c measure is the threshold c thre In addition, the measured value c measure Less than the threshold c thre In the case of , just invert the sign.

[0119] c=c measure -c thre (6)

[0120] In addition, when the constraint condition is that the measured value of the parameter needs to fall within a prescribed value range, the upper limit value of the prescribed range is set to c upper And set the lower limit to c lower When the constraint condition c represented by the following formula (7) is calculated 1 and the constraint condition c represented by the following equation (8) 2 In the second embodiment, the number of constraints is not limited. For example, when k is a natural number, the parameter search device 1A searches for the kth constraint c k The parameter can be 0 or more.

[0121] c 1 =c upper -c measure (7)

[0122] c 2 =c measure -c lower (8)

[0123] The operation result storage unit 112A stores the operation result acquired by the operation result acquisition unit 111A. The operation result storage unit 112A is a storage device provided in the computer functioning as the parameter search device 1A, and includes a storage device such as a HDD or SSD or a storage device such as a hard disk drive ... Figure 5B Memory 104, etc.

[0124] The evaluation value acquisition unit 12A acquires the evaluation value of the parameter obtained using the operation result and also acquires the constraint condition. The constraint condition calculation unit 123 uses the operation result obtained by the operation result acquisition unit 111A to calculate the constraint condition to be satisfied in the operation of the mechanical device 2. For example, the constraint condition calculation unit 123 uses the operation result read from the operation result storage unit 112A to calculate the constraint condition according to the above formula (6) or the above formula (7) and the above formula (8). The constraint condition calculated by the constraint condition calculation unit 123 is stored in the evaluation value storage unit 122A in correspondence with the parameter and the evaluation value.

[0125] The evaluation value storage unit 122A stores parameters, evaluation values, and constraints. The evaluation value storage unit 122A is a storage device provided in a computer functioning as the parameter search device 1, and includes a storage device such as a HDD or SSD or a storage device such as a hard disk drive ... Figure 5B Memory 104, etc.

[0126] The machine learning unit 131A generates a machine learning model for predicting the relationship between parameters, evaluation values, and constraints in the search candidates based on the operation results of the mechanical device 2, and uses the machine learning model to predict the relationship between parameters, evaluation values, and constraints. Figure 8 As shown, the machine learning unit 131A includes a learning prediction unit 1311A and a prediction result storage unit 1312A.

[0127] The learning prediction unit 1311A generates search candidates for parameters and constraints based on the operation results of the mechanical device 2, and uses the search candidates as learning data to cause the machine learning model to learn the relationship between the parameters, constraints, and evaluation values ​​included in the operation results. In addition, the learning prediction unit 1311A uses the machine learning model to predict the relationship between the parameters, constraints, and evaluation values ​​in the search candidates.

[0128] For example, the learning prediction unit 1311A learns the relationship between parameters and constraints through a machine learning model, and uses the machine learning model to predict the relationship between parameters that satisfy the constraints and evaluation values. By using the prediction results, the parameter search device 1A can search for parameters that satisfy the constraints and have high evaluation values.

[0129] In addition, the evaluation value taking the constraints into consideration may be calculated. For example, a penalty function may be calculated in which the value increases as the value is outside the range indicated by the constraints and is farther from a fixed threshold, a barrier function may be calculated in which the value increases as the value is within the range indicated by the constraints and is closer to a fixed threshold, or a term obtained by combining the values ​​of these functions may be added to the evaluation value function representing the evaluation value.

[0130] For example, the learning prediction unit 1311A may also use a Gaussian process regression model to calculate the prediction result of the relationship between the parameter and the evaluation value, that is, the average μ and standard deviation σ of the search candidate point, and calculate the prediction result of the relationship between the constraint condition of the search candidate point and the parameter, that is, the average μ k and standard deviation σ k , using these prediction results, the acquisition function is calculated in the same way as in implementation mode 1.

[0131] As an acquisition function that takes constraints into consideration, there is, for example, Expected Improvement with Constraints (hereinafter referred to as EIC).

[0132] EIC can be expressed by the function acq in the following equation (9): EIC express.

[0133] As in the first embodiment, the search parameter calculation unit 1331A calculates the acquisition function acq EIC The value of becomes the largest search candidate point, and the calculated search candidate point is output to the parameter instruction unit 1332A as the next search object.

[0134]

[0135] The prediction result storage unit 1312A stores the prediction results of the machine learning model. The prediction result storage unit 1312A is a storage device provided in the computer functioning as the parameter search device 1A, and includes a storage device such as a HDD or SSD or a storage device such as a hard disk drive ... Figure 5B Memory 104, etc.

[0136] The search end determination unit 132A determines the end of the parameter search based on the prediction result of the machine learning model. For example, the learning prediction unit 1311A predicts the relationship between the evaluation value and the constraint condition and the parameter using a Gaussian process regression model.

[0137] By using the average μ and standard deviation σ, which are the prediction results of the evaluation value using the Gaussian process regression model, and the average μ, which are the prediction results of the constraint condition using the Gaussian process regression model, k and standard deviation σ k , Best Search best , and the threshold c associated with the constraint k , the search candidate points that satisfy the constraints can be expressed by the following formula (10).

[0138] As in the first embodiment, the search end determination unit 132A determines that the search end exceeds the optimal search y. best The number of search candidates is used as the search end indicator.

[0139] (μ k -c k ) / σ k ≥0 (10)

[0140] For example, if the search end index is greater than the threshold, the search end determination unit 132A determines to continue the parameter search, and when the search end index is less than the threshold, the search end determination unit 132A determines to end the parameter search. When it is determined to continue the parameter search, the search end determination unit 132A reads the prediction result of this time from the prediction result storage unit 1312A and outputs it to the parameter determination unit 133A.

[0141] When the parameter determination unit 133A determines to continue the parameter search based on the prediction result of the machine learning model, it determines the next parameter to be searched (search object). Figure 8 As shown, the parameter determination unit 133A includes a search parameter calculation unit 1331A and a parameter instruction unit 1332A.

[0142] The search parameter calculation unit 1331A calculates the next parameter to be searched. For example, the search parameter calculation unit 1331A calculates the search candidate whose relationship among the parameters, evaluation values ​​and constraints in the current prediction result about the relationship among the parameters, evaluation values ​​and constraints is closest to the target as the next search object. In the case where Bayesian optimization is used for parameter search, the search candidate point whose value of the acquisition function considering the average μ and standard deviation σ of the search candidate points is the largest is the next search object.

[0143] The parameter instruction unit 1332A instructs the mechanical device 2 on the parameters calculated by the search parameter calculation unit 1331A. For example, the parameter instruction unit 1332A communicates with the mechanical device 2 via a wireless or wired communication device ( Figure 8 (not shown) sends the parameters of the search object to the mechanical device 2. The mechanical device 2 operates under the operating conditions indicated by the parameters received from the parameter instruction unit 1332A.

[0144] Furthermore, although the operation result storage unit 112A, the evaluation value storage unit 122A, and the prediction result storage unit 1312A are shown as separate storage units, these storage units may be provided in a storage area of ​​one storage device.

[0145] Furthermore, the operation result storage unit 112A, the evaluation value storage unit 122A, and the prediction result storage unit 1312A are storage units included in the parameter search device 1A, but they may be storage units included in an external storage device provided separately from the parameter search device 1A.

[0146] In addition, the external storage device is a storage device that can read out storage contents from the parameter search device 1A through wireless or wired communication.

[0147] As described above, in the parameter search device 1A of the second embodiment, the operation result collection unit 11A collects the operation results including the constraints to be satisfied in the operation of the mechanical device 2. The parameter search unit 13A predicts the relationship between the parameters, evaluation values, and constraints in the search candidates using a machine learning model, and when a plurality of search candidates regarding the parameters, constraints, and evaluation values ​​are input, the machine learning model outputs a predicted value indicating the relationship between the parameters, evaluation values, and constraints in the search candidates. Thus, the parameter search device 1A can efficiently search for parameters indicating the target operation conditions of the mechanical device 2 without determining a fixed number of searches or search time.

[0148] Implementation method 3.

[0149] In the third embodiment, a parameter indicating the operating condition of the mechanical device 2 is searched based on a simulator simulating the operation of the mechanical device 2 or data acquired in advance for the mechanical device 2 .

[0150] Fig. 9 is a block diagram showing the structure of a parameter search device 1B according to Embodiment 3. Fig. 9 In the embodiment, the parameter search device 1B uses the prior data collected by the data collection device 3 to make the machine learning model learn the relationship between the parameters, evaluation values, and constraints representing the operating conditions of the mechanical device 2. The parameter search device 1B searches for the target operating conditions of the mechanical device 2 using the prediction results of the relationship between the parameters, evaluation values, and constraints based on the machine learning model.

[0151] The advance data collection device 3 obtains data related to the characteristics of the mechanical device 2 before the parameter search device 1B searches for parameters. The advance data collection device 3 includes a advance data acquisition unit 31 and a advance data storage unit 32. Fig. 9 As shown, the prior data collection device 3 is an external device provided separately from the parameter search device 1B. Alternatively, the prior data collection device 3 may be an internal device of the parameter search device 1B.

[0152] The advance data acquisition unit 31 acquires data related to the characteristics of the mechanical device 2. For example, the data related to the characteristics of the mechanical device 2 is data used to predict the relationship between the parameter and the evaluation value or the relationship between the parameter and the constraint condition. The parameter search device 1B uses the data related to the characteristics of the mechanical device 2 as learning data to predict the relationship between the parameter and the evaluation value or the relationship between the parameter and the constraint condition.

[0153] For example, the advance data acquisition unit 31 acquires data related to the characteristics of the mechanical device 2 from an actual machine of the mechanical device 2 or a simulator that simulates the operation of the mechanical device 2. The advance data acquired by the advance data acquisition unit 31 is stored in the advance data storage unit 32.

[0154] The advance data storage unit 32 is a storage device that stores the advance data acquired by the advance data acquisition unit 31. The advance data storage unit 32 may be a storage device provided in the advance data collection device 3, or may be an external storage device provided separately from the advance data collection device 3. For example, the advance data storage unit 32 may be a cloud device that can read and write from the parameter search device 1B and the advance data acquisition unit 31 via a network.

[0155] The parameter search device 1B includes an operation result collection unit 11A, an evaluation value acquisition unit 12A, and a parameter search unit 13B. The operation result collection unit 11A includes an operation result acquisition unit 111A and an operation result storage unit 112A. The evaluation value acquisition unit 12A includes an evaluation value calculation unit 121, an evaluation value storage unit 122A, and a constraint calculation unit 123. The parameter search unit 13B includes a machine learning unit 131B, a search end determination unit 132B, and a parameter determination unit 133B.

[0156] The operation result acquisition unit 111A acquires an operation result including, in addition to parameters indicating the operation conditions of the mechanical device 2, constraints during operation of the mechanical device 2. The operation result storage unit 112A stores the operation result acquired by the operation result acquisition unit 111A.

[0157] The evaluation value acquisition unit 12A acquires the evaluation value of the parameter obtained using the operation result and also acquires the constraint condition. The constraint condition calculation unit 123 calculates the constraint condition to be satisfied in the operation of the mechanical device 2 using the operation result obtained by the operation result acquisition unit 111A.

[0158] The evaluation value storage unit 122A stores parameters, evaluation values, and constraints. The evaluation value storage unit 122A is a storage device provided in a computer functioning as the parameter search device 1, and includes a storage device such as a HDD or SSD or a storage device such as a hard disk drive ... Figure 5B Memory 104, etc.

[0159] The machine learning unit 131B generates a machine learning model for predicting the relationship between parameters, evaluation values, and constraints in search candidates based on the operation results of the mechanical device 2, and uses the machine learning model to predict the relationship between parameters, evaluation values, and constraints.

[0160] like Fig. 9 As shown, the machine learning unit 131B includes a learning prediction unit 1311B, a prediction result storage unit 1312B, and a prior learning unit 1313.

[0161] The learning prediction unit 1311B generates search candidate points of parameters and constraints based on the operation results of the mechanical device 2, and uses the search candidate points as learning data to make the machine learning model learn the relationship between the parameters, constraints and evaluation values. Furthermore, the learning prediction unit 1311B generates the above-mentioned machine learning model using the prior learning model learned by the prior learning unit 1313.

[0162] For example, the above-mentioned machine learning model and the prior learning model are the same type of models. The learning prediction unit 1311B uses the parameters, evaluation values, and constraints stored in the evaluation value storage unit 122A to further learn the relationship between the parameters, evaluation values, and constraints through the prior learning model, thereby using the learned prior learning model as the above-mentioned machine learning model.

[0163] The parameter search device 1B uses a previously learned model to efficiently search for parameters representing the operating conditions of the mechanical device 2 even if the amount of data representing the operating results of the mechanical device 2 is small.

[0164] The learning prediction unit 1311B predicts the relationship between parameters, constraints, and evaluation values ​​in the search candidates using a machine learning model.

[0165] For example, the learning prediction unit 1311B learns the relationship between parameters and constraints through a machine learning model, and uses the machine learning model to predict the relationship between parameters that satisfy the constraints and evaluation values. By using the prediction results, the parameter search device 1B can search for parameters that satisfy the constraints and have high evaluation values.

[0166] The prediction result storage unit 1312B stores the prediction results of the machine learning model. The prediction result storage unit 1312B is a storage device provided in the computer functioning as the parameter search device 1B, including a storage device such as a HDD or SSD or Figure 5B Memory 104, etc.

[0167] The prior learning unit 1313 learns the relationship between the parameters, evaluation values, and constraints in the search candidates through a prior learning model using the prior data stored in the prior data storage unit 32. The prior learning model that has learned the relationship between the parameters, evaluation values, and constraints is output to the learning prediction unit 1311B.

[0168] The search end determination unit 132B determines the end of the parameter search based on the prediction result of the machine learning model. For example, the learning prediction unit 1311B uses a Gaussian process regression model to predict the relationship between the evaluation value and the constraint condition and the parameter. For example, if the search end index is above the threshold, the search end determination unit 132B determines to continue the parameter search. When the search end index is less than the threshold, the search end determination unit 132B determines to end the parameter search. When it is determined to continue the parameter search, the search end determination unit 132B reads the prediction result of this time from the prediction result storage unit 1312B and outputs it to the parameter determination unit 133B.

[0169] When the parameter determination unit 133B determines to continue the parameter search based on the prediction result of the machine learning model, it determines the next parameter to be searched. Fig. 9 As shown, the parameter determination unit 133B includes a search parameter calculation unit 1331B and a parameter instruction unit 1332B.

[0170] The search parameter calculation unit 1331B calculates the next parameter to be searched.

[0171] For example, the search parameter calculation unit 1331B calculates the search candidate whose relationship among parameters, evaluation values, and constraints is closest to the target among the current prediction results regarding the relationship among parameters, evaluation values, and constraints as the next search object.

[0172] When Bayesian optimization is used for parameter search, the search parameter calculation unit 1331B sets the search candidate having the largest value of the acquisition function taking into account the average μ and standard deviation σ of the search candidates as the next search target.

[0173] The parameter instruction unit 1332B instructs the mechanical device 2 on the parameters calculated by the search parameter calculation unit 1331B. For example, the parameter instruction unit 1332B communicates with the mechanical device 2 via a wireless or wired communication device ( Fig. 9 (not shown) sends the parameters of the search object to the mechanical device 2. The mechanical device 2 operates according to the operating conditions indicated by the parameters received from the parameter instruction unit 1332B.

[0174] Furthermore, although the operation result storage unit 112A, the evaluation value storage unit 122A, and the prediction result storage unit 1312B are shown as separate storage units, these storage units may be provided in a storage area of ​​one storage device.

[0175] Furthermore, the operation result storage unit 112A, the evaluation value storage unit 122A, and the prediction result storage unit 1312B are storage units included in the parameter search device 1B, but they may be storage units included in an external storage device provided separately from the parameter search device 1B.

[0176] In addition, the external storage device is a storage device capable of reading out storage contents from the parameter search device 1B through wireless or wired communication.

[0177] Fig.10 It is a flowchart showing the parameter search method according to the third embodiment, and shows the operations of the parameter search device 1B, the mechanical device 2 , and the advance data collection device 3 .

[0178] First, the advance data collection device 3 confirms whether the advance data exists in a predetermined storage location (step ST1B). For example, the predetermined storage location includes a cloud on the network.

[0179] If the prior data exists (step ST1B; yes), the prior data collection device 3 accesses the storage location and collects the prior data from the storage location (step ST2B). For example, the prior data acquisition unit 31 communicates with the cloud on the network and downloads the prior data. The prior data acquired by the prior data acquisition unit 31 is stored in the prior data storage unit 32.

[0180] Next, the advance learning unit 1313 uses the advance data read from the advance data storage unit 32 as learning data, and learns the relationship between the parameters, evaluation values, and constraints that represent the operating conditions of the mechanical device 2 through the advance learning model (step ST3B). The advance learning model that has learned the relationship between the parameters, evaluation values, and constraints is output to the learning prediction unit 1311B.

[0181] When there is no prior data in the storage location (step ST1B; No), or when a prior learning model is generated in step ST3B, the search parameter calculation unit 1331B determines the initial value (initial point) of the parameter to be searched (step ST4B). For example, the search parameter calculation unit 1331B randomly determines the initial value of the parameter to be searched from the parameters representing various operating conditions of the mechanical device 2. In addition, when a parameter with a high evaluation value is known in advance, the parameter may be set as the initial value.

[0182] Next, the parameter instruction unit 1332B instructs the mechanical device 2 to set the initial value of the search target parameter calculated by the search parameter calculation unit 1331B (step ST5B). For example, the parameter instruction unit 1332B sends the initial value of the search target parameter to the mechanical device 2 via a wireless or wired communication device.

[0183] When the mechanical device 2 receives the initial value of the parameter from the parameter instruction unit 1332B, it operates according to the operating conditions indicated by the received parameter value (step ST6B). When the mechanical device 2 starts to operate, the operation result acquisition unit 111A sequentially acquires the operation results from the mechanical device 2 (step ST7B). The operation results acquired by the operation result acquisition unit 111A are stored in the operation result storage unit 112A.

[0184] The evaluation value calculation unit 121A calculates the evaluation value of the parameter included in the operation result stored in the operation result storage unit 112A. The evaluation value calculated by the evaluation value calculation unit 121A is stored in the evaluation value storage unit 122A in correspondence with the parameter. The constraint condition calculation unit 123 calculates the constraint condition using the operation result stored in the operation result storage unit 112A. The constraint condition calculated by the constraint condition calculation unit 123 is stored in the evaluation value storage unit 122A in correspondence with the parameter.

[0185] The learning prediction unit 1311B uses the parameters, evaluation values, and constraints stored in the evaluation value storage unit 122A as learning data, and causes the machine learning model (or pre-learning model) to learn the relationship between the parameters, evaluation values, and constraints (step ST8B). The learning prediction unit 1311B generates search candidate points for the parameters representing the operating conditions of the mechanical device 2 (step ST9B).

[0186] Next, the learning prediction unit 1311B uses the machine learning model (or the pre-learned model) to predict the relationship between the parameters, evaluation values, and constraints of the search candidate points (step ST10B). For example, when Bayesian optimization is used to determine the parameters of the search object, the learning prediction unit 1311B calculates the function value called the acquisition function and uses the acquisition function value to calculate the next search object.

[0187] The search end determination unit 132B determines whether to end the parameter search based on the prediction results stored in the prediction result storage unit 1312B (step ST11B). If it is determined that the search end index is less than the threshold value (step ST11B; yes), the search end determination unit 132B determines to end the parameter search. Fig.10 On the other hand, when it is determined that the search end indicator is above the threshold (step ST11B; No), the search end determination unit 132B determines to continue the parameter search, reads the prediction result of this time from the prediction result storage unit 1312B and outputs it to the parameter determination unit 133B.

[0188] The search parameter calculation unit 1331B calculates the search candidate point with the largest value of the acquisition function taking into account the average μ and standard deviation σ of the search candidate points as the next search object (step ST12B). The parameter instruction unit 1332B indicates the parameters of the search object calculated by the search parameter calculation unit 1331B to the mechanical device 2. The mechanical device 2 operates according to the operating conditions indicated by the parameters received from the parameter instruction unit 1332B.

[0189] After that, the process returns to step ST5B and the above-mentioned process is repeated.

[0190] Furthermore, the parameter search device 1B according to the third embodiment is obtained by adding the pre-learning unit 1313 to the configuration of the second embodiment, but the pre-learning unit 1313 may be added to the configuration of the first embodiment.

[0191] For example, the prior learning unit 1313 uses the prior data stored in the prior data storage unit 32 to learn the relationship between the parameters and the evaluation values ​​in the search candidates through the prior learning model. The prior learning model that has learned the relationship between the parameters and the evaluation values ​​is output to the learning prediction unit 1311B. The learning prediction unit 1311B uses the parameters and evaluation values ​​stored in the evaluation value storage unit 122A to further learn the relationship between the parameters and the evaluation values ​​through the prior learning model, thereby using the learned prior learning model as the above-mentioned machine learning model.

[0192] The parameter search device 1B uses a previously learned model to efficiently search for parameters representing the operating conditions of the mechanical device 2 even if the amount of data representing the operating results of the mechanical device 2 is small.

[0193] As described above, in the parameter search device 1B according to the third embodiment, the parameter search unit 13B acquires the pre-learned model that has previously learned the relationship between the parameter and the evaluation value, and learns the machine learning model using the pre-learned model.

[0194] When the machine learning model has less learning data, the learning becomes insufficient and it is not possible to make predictions with high accuracy. Therefore, the efficiency of parameter search using the machine learning model with insufficient learning also deteriorates. In contrast, by using the model learned in advance, the parameter search device 1B can efficiently search for parameters representing the operating conditions of the mechanical device 2 even if the data representing the operating results of the mechanical device 2 is small.

[0195] In the parameter search device 1B according to the third embodiment, the parameter search unit 13B obtains a pre-learned model that has previously learned the relationship between parameters, constraints, and evaluation values, and trains the machine learning model using the pre-learned model.

[0196] By using the previously learned model, the parameter search device 1B can efficiently search for parameters representing the operating conditions of the mechanical device 2 even if the data representing the operating results of the mechanical device 2 is small.

[0197] Implementation method 4.

[0198] In addition to using the operating results of the mechanical device, embodiment 4 also uses external information detected by external sensors installed in the mechanical device to predict the relationship between parameters representing the operating conditions of the mechanical device, evaluation values ​​and external information, or the relationship between parameters, constraints and external information.

[0199] Fig.11 1C is a block diagram showing the structure of a parameter search device 1C according to a fourth embodiment. Fig.11 In the embodiment, the parameter search device 1C uses the prior data collected by the prior data collection device 3 and the external information detected by the external sensor 4 to learn the relationship between the parameters, evaluation values, and constraints representing the operating conditions of the mechanical device 2. The parameter search device 1C searches for the target operating conditions of the mechanical device 2 using the predicted results of the relationship between the parameters, evaluation values, and constraints.

[0200] The advance data collection device 3 obtains data related to the characteristics of the mechanical device 2 before the parameter search device 1C searches for parameters. The advance data collection device 3 includes a advance data acquisition unit 31 and a advance data storage unit 32. Fig.11 As shown, the prior data collection device 3 is an external device provided separately from the parameter search device 1C.

[0201] In addition, the advance data collection device 3 may be an internal device of the parameter search device 1C.

[0202] The prior data acquisition unit 31 acquires data related to the characteristics of the mechanical device 2. For example, the data related to the characteristics of the mechanical device 2 is data used to predict the relationship between the parameter and the evaluation value or the relationship between the parameter and the constraint condition. The parameter search device 1C uses the data related to the characteristics of the mechanical device 2 as learning data to predict the relationship between the parameter and the evaluation value or the relationship between the parameter and the constraint condition.

[0203] The advance data storage unit 32 is a storage device that stores the advance data acquired by the advance data acquisition unit 31. The advance data storage unit 32 may be a storage device provided in the advance data collection device 3, or may be an external storage device provided separately from the advance data collection device 3. For example, the advance data storage unit 32 may be a cloud device that can read and write from the parameter search device 1C and the advance data acquisition unit 31 via a network.

[0204] The external sensor 4 is a sensor provided around or inside the mechanical device 2, and detects external information about the mechanical device 2. The external information is information including at least one of the peripheral information of the mechanical device 2 and information related to the operation task. For example, when the mechanical device 2 is a processing machine, the external information is information indicating the temperature and vibration of the mechanical device 2 or the shape of the processed workpiece. These are information that affects the smoothness and processing speed of the processed shape of the workpiece.

[0205] The parameter search device 1C includes an operation result collection unit 11B, an evaluation value acquisition unit 12B, and a parameter search unit 13C. The operation result collection unit 11B includes an operation result acquisition unit 111B, an operation result storage unit 112B, a sensor data acquisition unit 113, and a sensor data storage unit 114. The evaluation value acquisition unit 12B includes an evaluation value calculation unit 121B, an evaluation value storage unit 122B, and a constraint calculation unit 123A. The parameter search unit 13C includes a machine learning unit 131C, a search end determination unit 132C, and a parameter determination unit 133C.

[0206] The operation result acquisition unit 111A acquires the operation result including the constraint conditions during the operation of the mechanical device 2 in addition to the parameters indicating the operation conditions of the mechanical device 2. The operation result storage unit 112A stores the operation result acquired by the operation result acquisition unit 111A.

[0207] The sensor data acquisition unit 113 acquires external information from the external sensor 4. For example, the sensor data acquisition unit 113 acquires external information from the external sensor 4 through wired or wireless communication, and stores the acquired external information in the sensor data storage unit 114. The sensor data storage unit 114 stores the external information acquired by the sensor data acquisition unit 113.

[0208] The evaluation value acquisition unit 12B acquires the evaluation value of the parameter obtained using the operation result of the mechanical device 2 and the external information, and also acquires the constraint condition. The evaluation value calculation unit 121B calculates the evaluation value of the parameter using the operation result, and calculates the evaluation value of the parameter using the external information. Similarly, the constraint condition calculation unit 123A calculates the constraint condition to be satisfied in the operation of the mechanical device 2 using the operation result, and calculates the constraint condition using the external information.

[0209] The evaluation value storage unit 122B stores parameters, evaluation values, and constraints. The evaluation value storage unit 122B is a storage device provided in a computer functioning as the parameter search device 1, and includes a storage device such as a HDD or SSD or a storage device such as a hard disk drive ... Figure 5B Memory 104, etc.

[0210] The machine learning unit 131C generates a machine learning model for predicting the relationship between parameters, evaluation values, and constraints in search candidates based on the operation results of the mechanical device 2, and predicts the relationship between parameters, evaluation values, and constraints using the machine learning model.

[0211] like Fig.11 As shown, the machine learning unit 131C includes a learning prediction unit 1311C, a prediction result storage unit 1312C, and a prior learning unit 1313.

[0212] The learning prediction unit 1311C generates search candidates for parameters and constraints based on the operation results of the mechanical device 2, and uses the search candidates as learning data to enable the machine learning model to learn the relationship between the parameters, constraints, and evaluation values ​​included in the operation results.

[0213] In addition, the learning prediction unit 1311C generates the above-mentioned machine learning model using the pre-learning model learned by the pre-learning unit 1313.

[0214] For example, the machine learning model and the prior learning model are the same type of models. The learning prediction unit 1311C uses the parameters, evaluation values, and constraints stored in the evaluation value storage unit 122C to further learn the relationship between the parameters, evaluation values, and constraints through the prior learning model, thereby using the learned prior learning model as the machine learning model.

[0215] The parameter search device 1C can efficiently search for parameters representing the operating conditions of the mechanical device 2 even if the data representing the operating results of the mechanical device 2 is small, by using the previously learned model.

[0216] The learning prediction unit 1311C uses a machine learning model to predict the relationship between parameters, constraints, and evaluation values ​​in search candidates.

[0217] For example, the learning prediction unit 1311C learns the relationship between parameters and constraints through a machine learning model, and uses the machine learning model to predict the relationship between parameters that satisfy the constraints and the evaluation value. By using the prediction results, the parameter search device 1C can search for parameters that satisfy the constraints and have a high evaluation value.

[0218] The prediction result storage unit 1312C stores the prediction results of the machine learning model. The prediction result storage unit 1312C is a storage device provided in the computer functioning as the parameter search device 1C, including a storage device such as a HDD or SSD or Figure 5B Memory 104, etc.

[0219] The prior learning unit 1313 learns the relationship between parameters, evaluation values, and constraints in the search candidates through a prior learning model using the prior data stored in the prior data storage unit 32. The prior learning model that has learned the relationship between parameters, evaluation values, and constraints is output to the learning prediction unit 1311C.

[0220] The search end determination unit 132C determines the end of the parameter search based on the prediction result of the machine learning model. For example, the learning prediction unit 1311C uses a Gaussian process regression model to predict the relationship between the evaluation value and the constraint condition and the parameter. For example, if the search end index is above the threshold, the search end determination unit 132C determines to continue the parameter search. When the search end index is less than the threshold, the search end determination unit 132C determines to end the parameter search. When it is determined to continue the parameter search, the search end determination unit 132C reads the prediction result of this time from the prediction result storage unit 1312C and outputs it to the parameter determination unit 133C.

[0221] When the parameter determination unit 133C determines to continue the parameter search based on the prediction result of the machine learning model, it determines the next parameter to be searched. Fig.11 As shown, the parameter determination unit 133C includes a search parameter calculation unit 1331C and a parameter instruction unit 1332C.

[0222] The search parameter calculation unit 1331C calculates the next parameter to be searched.

[0223] For example, the search parameter calculation unit 1331C calculates the search candidate point whose relationship among parameters, evaluation values ​​and constraints is closest to the target among the current prediction results regarding the relationship among parameters, evaluation values ​​and constraints as the next search object.

[0224] When Bayesian optimization is used for parameter search, the search parameter calculation unit 1331C sets the search candidate point having the largest value of the acquisition function taking into account the average μ and standard deviation σ of the search candidates as the next search target.

[0225] The parameter instruction unit 1332C instructs the mechanical device 2 on the parameters calculated by the search parameter calculation unit 1331C. For example, the parameter instruction unit 1332C communicates with the mechanical device 2 via a wireless or wired communication device ( Fig.11 (not shown) the parameters of the search object are sent to the mechanical device 2. The mechanical device 2 operates according to the operating conditions indicated by the parameters received from the parameter instruction unit 1332C.

[0226] Furthermore, although the operation result storage unit 112A, the evaluation value storage unit 122B, and the prediction result storage unit 1312C are shown as separate storage units, these storage units may be provided in a storage area of ​​one storage device.

[0227] Furthermore, the operation result storage unit 112A, the evaluation value storage unit 122B, and the prediction result storage unit 1312C are storage units included in the parameter search device 1C, but they may be storage units included in an external storage device provided separately from the parameter search device 1C.

[0228] In addition, the external storage device is a storage device capable of reading out storage contents from the parameter search device 1C through wireless or wired communication.

[0229] In addition, the parameter search device 1C of the fourth embodiment is obtained by replacing the operation result collection unit 11A in the configuration of the third embodiment with the operation result collection unit 11B and replacing the evaluation value acquisition unit 12A with the evaluation value acquisition unit 12B.

[0230] In addition, the parameter search device 1C may replace only the operation result collection unit 11 in the parameter search device 1 of the first embodiment with the operation result collection unit 11B, and output only the external information from the sensor data storage unit 114 to the evaluation value calculation unit 121 .

[0231] In this case, the evaluation value calculation unit 121 calculates the evaluation value of the parameter using the operation result and the evaluation value of the parameter using the external information. Since the machine learning model can be learned using the peripheral information of the mechanical device 2 or the information related to the operation task, the prediction accuracy of the machine learning model is improved. As a result, the parameter search device 1C can efficiently search for parameters representing the operating conditions of the mechanical device 2.

[0232] As described above, in the parameter search device 1C of the fourth embodiment, the operation result collection unit 11B further obtains external information including at least one of the peripheral information of the mechanical device 2 and the information related to the operation task. The evaluation value acquisition unit 12B uses the external information to calculate the evaluation value of the parameter. Since the machine learning model can be learned using the external information, the prediction accuracy of the relationship between the parameter and the evaluation value based on the machine learning model is improved. As a result, the parameter search device 1C can efficiently search for parameters representing the operating conditions of the mechanical device 2.

[0233] In the parameter search device 1C of the fourth embodiment, the operation result collection unit 11B further obtains external information including at least one of the peripheral information of the mechanical device 2 and the information related to the operation task. The evaluation value acquisition unit 12B uses the external information to calculate the evaluation value of the constraint condition. Since the machine learning model can be learned using the external information, the relationship prediction accuracy among the parameters, evaluation values, and constraint conditions based on the machine learning model is improved. As a result, the parameter search device 1C can efficiently search for parameters representing the operating conditions of the mechanical device 2.

[0234] Implementation method 5.

[0235] Fig.12 : is a block diagram showing the structure of parameter search device 1 of Embodiment 5. In Embodiment 5, the mechanical device is air conditioning and heating equipment 2A, and parameter search device 1 of Embodiment 1 searches for parameters indicating operating conditions of air conditioning and heating equipment 2A.

[0236] Fig.13 2A is a conceptual diagram showing a refrigeration cycle of an air conditioning and cooling device 2A. The air conditioning and cooling device 2A is, for example, an air conditioner, a ventilation device, a sanitary device, a refrigerator, or a water heater. Fig.13 As shown, the air-conditioning cooling and heating device 2A includes an indoor unit 21 and an outdoor unit 22 .

[0237] The indoor unit 21 includes an indoor heat exchanger 211 and an electronic expansion valve 212. The outdoor unit 22 includes an electronic expansion valve 221 for bypass, a refrigerant-refrigerant heat exchanger 222, a four-way valve 223, a compressor 224, a liquid storage tank 225, an outdoor heat exchanger 226, a solenoid valve 227A, and a solenoid valve 227B. The air conditioning cooling and heating device 2A is composed of a compressor 224 for conveying refrigerant, a condenser for releasing heat of the refrigerant to the surrounding fluid, an evaporator for absorbing heat from the surrounding fluid to the refrigerant, and an electronic expansion valve 212 and an electronic expansion valve 221 for applying a pressure difference to the refrigerant. The condenser is the outdoor heat exchanger 226 of the outdoor unit 22, and the evaporator is the indoor heat exchanger 211 of the indoor unit 21.

[0238] The operation result collecting unit 11 collects operation results including parameters indicating the operation conditions of the air conditioning and cooling equipment 2A. For example, the parameters include the opening of the electronic expansion valve 212, the opening of the electronic expansion valve 221, the opening of the four-way valve 223, the opening of the electromagnetic valves 227A and 227B, the operating frequency of the compressor 224, the air volume of the fan, the speed of the fan, the angle of the blades that determine the blowing direction, and at least one of the flow rate of the water to be controlled for heating and cooling. Thus, the parameter search device 1 can efficiently search for parameters indicating the operation conditions of the air conditioning and cooling equipment 2A.

[0239] The evaluation value calculation unit 121 uses the operation result of the air conditioning device 2A to calculate the evaluation value of the parameter representing the operation condition of the air conditioning device 2A. For example, the evaluation value includes the cooling and heating capacity of the air conditioning device 2A, the energy efficiency (COP) representing the cooling and heating capacity corresponding to the power consumption of the air conditioning device 2A, the comfort index (PMV) of the air conditioning device 2A, the temperature of the air outlet of the air conditioning device 2A, the outlet temperature of the water to be controlled by the air conditioning device 2A, and at least one of the carbon dioxide concentration discharged by the air conditioning device 2A. Thus, the parameter search device 1 can efficiently search for the parameters representing the operation condition of the air conditioning device 2A.

[0240] For example, when the COP is to be maximized in rated cooling, the COP cannot be directly measured. Therefore, the COP is calculated by the evaluation value calculation unit 121 according to the following equation (11). The parameter search unit 13 searches for a parameter that maximizes the COP value.

[0241] Furthermore, the evaluation value calculation unit 121 can usually calculate the cooling and heating capabilities based on the heat balance in the indoor unit 21 (the difference in temperature or humidity of the sucked or blown air and the air volume).

[0242] Alternatively, the learning prediction unit 1311 may calculate the physical properties of the refrigerant using the accumulator 225 or the indoor and outdoor air temperatures, and predict the heat balance of the refrigerant based on the calculation results.

[0243] COP = rated capacity (kW) / rated power consumption (kW) (11)

[0244] Fig.14 1 is a flowchart showing a parameter search method according to Embodiment 5, and shows the operations of the parameter search device 1 and the air conditioning cooling and heating equipment 2A.

[0245] First, the search parameter calculation unit 1331 determines the initial value of the parameter of the search object (step ST1C). For example, the search parameter calculation unit 1331 randomly determines the initial value of the parameter of the search object from the parameters representing various operating conditions of the air conditioning cooling and heating device 2A. In addition, if a parameter with a high evaluation value is known in advance, the parameter can also be set as the initial value.

[0246] Next, the parameter instruction unit 1332 indicates to the air conditioning cooling and heating device 2A the initial value of the parameter of the search object calculated by the search parameter calculation unit 1331 (step ST2C). For example, the parameter instruction unit 1332 sends the initial value of the parameter of the search object calculated by the search parameter calculation unit 1331 to the air conditioning cooling and heating device 2A via the air conditioning controller ( Fig.12 (not shown in the figure), the initial value of the parameter of the search object is sent to the air conditioning heating and cooling equipment 2A for setting.

[0247] When receiving the initial value of the parameter from the parameter instruction unit 1332, the air conditioning cooling and heating device 2A operates according to the operating conditions indicated by the received parameter value (step ST3C).

[0248] When the air conditioning and cooling device 2A starts operating, the operation result acquisition unit 111 sequentially acquires the operation results from the air conditioning and cooling device 2A via the air conditioning controller (step ST4C). The operation results acquired by the operation result acquisition unit 111 are stored in the operation result storage unit 112.

[0249] Next, the operation result acquisition unit 111 confirms whether the cooling capacity of the air-conditioning cooling and heating device 2A is predicted (step ST5C). For example, the operation result acquisition unit 111 displays a confirmation screen on the air-conditioning controller and inquires the user.

[0250] When the cooling capacity of the air-conditioning and cooling device 2A is predicted (step ST5C; Yes), the operation result acquisition unit 111 predicts the cooling capacity of the air-conditioning and cooling device 2A (step ST6C).

[0251] For example, the operation result acquisition unit 111 uses the liquid storage device of the air conditioning device 2A or the indoor and outdoor air temperatures to predict the heat balance of the refrigerant of the air conditioning device 2A, and uses the prediction result of the heat balance of the refrigerant to calculate the cooling capacity of the air conditioning device 2A. Next, the evaluation value calculation unit 121 calculates the evaluation value of the parameter representing the cooling capacity of the air conditioning device 2A. The evaluation value calculated by the evaluation value calculation unit 121 is stored in the evaluation value storage unit 122 in correspondence with the parameter.

[0252] When the cooling capacity of the air conditioning and cooling device 2A is not predicted (step ST5C; No), the operation result acquisition unit 111 calculates a parameter indicating the cooling capacity of the air conditioning and cooling device 2A using the operation results stored in the operation result storage unit 112 (step ST7C).

[0253] For example, the operation result acquisition unit 111 calculates the cooling capacity of the air conditioning device 2A using the difference in temperature or humidity of the air at the air inlet or outlet of the air conditioning device 2A and the measured data of the air volume. The evaluation value calculation unit 121 calculates the evaluation value of the parameter representing the cooling capacity of the air conditioning device 2A. The evaluation value calculated by the evaluation value calculation unit 121 is stored in the evaluation value storage unit 122 in correspondence with the parameter.

[0254] The learning prediction unit 1311 uses the parameters and evaluation values ​​stored in the evaluation value storage unit 122 as learning data, and causes the machine learning model to learn the relationship between the parameters and the evaluation values ​​(step ST8C). Next, the learning prediction unit 1311 generates search candidates for parameters representing the cooling capacity of the air conditioning cooling and heating device 2A (step ST9C). For example, the learning prediction unit 1311 uses the information representing the parameters and evaluation values ​​stored in the evaluation value storage unit 122 to randomly generate a plurality of search candidates.

[0255] The learning prediction unit 1311 uses the machine learning model to predict the relationship between the parameters and the evaluation value in the search candidate (step ST10C). For example, when Bayesian optimization is used to determine the parameters of the search object, the learning prediction unit 1311 calculates the next search object by calculating the function value called the acquisition function.

[0256] The search end determination unit 132 determines whether to end the parameter search based on the prediction results stored in the prediction result storage unit 1312 (step ST11C).

[0257] For example, the search end determination unit 132 calculates the prediction results of the Gaussian process regression model as a machine learning model, that is, the average μ, standard deviation σ, and optimal search y of the search candidates. best The comparison will exceed the optimal search y best The number of search candidates is used as the search end indicator S t .

[0258] When the search end indicator S is determined t If the value is less than the threshold value (step ST11C; Yes), the search end determination unit 132 determines to end the parameter search. Fig.14 On the other hand, when it is determined that the search end indicator S t If it is equal to or greater than the threshold value (step ST11C; No), the search end determination unit 132 determines to continue the parameter search, reads the current prediction result from the prediction result storage unit 1312 , and outputs it to the parameter determination unit 133 .

[0259] The search parameter calculation unit 1331 calculates the search candidate with the largest value of the acquisition function taking into account the average μ and standard deviation σ of the search candidates as the next search object (step ST12C). The parameter instruction unit 1332 indicates the parameters of the search object calculated by the search parameter calculation unit 1331 to the air conditioning and cooling equipment 2A. The air conditioning and cooling equipment 2A operates according to the operating conditions indicated by the parameters received from the parameter instruction unit 1332. After that, the process returns to step ST2C and the above process is repeated.

[0260] Fig.151 is a block diagram showing the structure of an air conditioning and heating device 2B having a parameter search device 1. Fig.15 As shown, the air conditioning and heating device 2B as a mechanical device may also include the parameter search device 1. The air conditioning and heating device 2B includes a flow meter 23 and an air conditioning controller 24 in addition to the parameter search device 1.

[0261] The flow meter 23 measures the air volume at the air inlet or outlet of the air conditioning device 2B. When the parameters of the search result are input, the air conditioning controller 24 sets them in the air conditioning device 2B. Thus, the air conditioning device 2B operates under the operating conditions indicated by the parameters searched by the parameter search device 1.

[0262] In Embodiment 5, parameter search device 1 searches for parameters indicating operating conditions of air conditioning and heating equipment 2A. Similarly, parameter search device 1A, 1B, or 1C may also search for parameters indicating operating conditions of air conditioning and heating equipment 2A.

[0263] When the parameter search device 1C searches for parameters of the air conditioning cooling and heating device 2A, the external information includes, for example, at least one of the dry-bulb temperature, wet-bulb temperature, gas pipe length, liquid pipe length, refrigerant amount, and heat load of the air conditioning cooling and heating device 2A. Thus, the parameter search device 1C can efficiently search for parameters representing the operating conditions of the air conditioning cooling and heating device 2A.

[0264] As described above, in the parameter search device 1 of the fifth embodiment, the mechanical device is the air conditioning cooling and heating device 2A. The parameters include at least one of the opening of the electronic expansion valve 212, the opening of the electronic expansion valve 221, the opening of the four-way valve 223, the opening of the electromagnetic valve 227A, the opening of the electromagnetic valve 227B, the operating frequency of the compressor 224, the air volume of the fan, the speed of the fan, the angle of the blades that determine the blowing direction, and the flow rate of the water to be controlled for heating and cooling. Thus, the parameter search device 1 can efficiently search for parameters representing the operating conditions of the air conditioning cooling and heating device 2A.

[0265] In the parameter search device 1 of the fifth embodiment, the evaluation value includes at least one of the cooling and heating capacity of the air conditioning device 2A, the COP indicating the cooling and heating capacity corresponding to the power consumption of the air conditioning device 2A, the PMV of the air conditioning device 2A, the temperature of the air outlet of the air conditioning device 2A, the outlet temperature of the water to be cooled and heated by the air conditioning device 2A, and the concentration of carbon dioxide discharged by the cooling and heating control of the air conditioning device 2A. Thus, the parameter search device 1 can efficiently search for parameters indicating the operating conditions of the air conditioning device 2A.

[0266] In the parameter search device 1C of the fifth embodiment, the external information includes at least one of the dry-bulb temperature, wet-bulb temperature, gas pipe length, liquid pipe length, refrigerant amount, and heat load of the air conditioning cooling and heating device 2A. Thus, the parameter search device 1C can efficiently search for parameters representing the operating conditions of the air conditioning cooling and heating device 2A.

[0267] Implementation method 6.

[0268] Fig.16 This is a block diagram showing the structure of a parameter search device 1A according to the sixth embodiment.

[0269] In the sixth embodiment, the mechanical device is an air conditioning and cooling device 2C, and the parameter search device 1A of the second embodiment searches for parameters indicating operating conditions of the air conditioning and cooling device 2C.

[0270] The constraint condition calculation unit 123 calculates the constraint condition to be satisfied in the operation of the air-conditioning and cooling device 2C, using the operation result acquired by the operation result acquisition unit 111A.

[0271] For example, when searching for parameters indicating operating conditions of the air conditioning device 2C that conforms to standards such as JRA (Japan Refrigeration and Air Conditioning Industry Association Standards), it is necessary to consider the stability of the refrigeration cycle of the air conditioning device 2C.

[0272] As the constraint conditions, for example, at least one of the cooling and heating capacity of the air conditioning and cooling device 2C, the power consumption of the air conditioning and cooling device 2C, the COP indicating the cooling and heating capacity corresponding to the power consumption of the air conditioning and cooling device 2C, the suction heating degree of the compressor of the air conditioning and cooling device 2C, the supercooling degree of the inlet of the electronic expansion valve of the air conditioning and cooling device 2C, the temperature and pressure of the refrigerant controlled by the compressor, the flow rate, temperature and pressure of the water to be cooled and heated controlled by the air conditioning and cooling device 2C, and the current flowing in the electronic circuit of the electronic substrate of the air conditioning and cooling device 2C is included. By considering these constraint conditions, the parameter search device 1A can efficiently search for parameters indicating the operating conditions of the air conditioning and cooling device 2C.

[0273] For example, if the constraint condition is that the cooling capacity is 40 kW or more, the constraint condition calculation unit 123 calculates the value of the constraint condition c using the following equation (12).

[0274] Constraint condition c = Refrigeration capacity measured value - 40 (kW) ≥ 0 (12)

[0275] In Embodiment 6, the parameter search device 1A searches for parameters indicating the operating conditions of the air conditioning and heating device 2C. Similarly, the parameter search device 1B or 1C may search for parameters indicating the operating conditions of the air conditioning and heating device 2A.

[0276] When the parameter search device 1B or 1C searches for parameters indicating the operating conditions of the air conditioning and cooling device 2C, the prior data acquired by the prior data collection device 3 is at least one of a plurality of parameters as search candidates and an evaluation value of the parameter.

[0277] Here, the advance data includes, for example, operation results obtained by simulating the operation of the air-conditioning and cooling device 2C using a simulator or log data when the air-conditioning and cooling device 2C was experimentally operated in the past.

[0278] When the state quantity and constraint conditions of the air conditioning and cooling equipment 2C can be obtained, the pre-learning unit 1313 uses the parameters representing the operating conditions of the air conditioning and cooling equipment 2C, the state quantity of the air conditioning and cooling equipment 2C, and the data representing the constraint conditions to generate a pre-learning model that has learned the relationship between the parameters and the constraint conditions. The pre-learning unit 1313 uses the pre-learning model to predict the relationship between the parameters and the constraint conditions and generates a pre-distribution that represents the distribution of the parameters and the constraint conditions. The learning prediction unit 1311B or 1311C uses the pre-distribution to make the machine learning model learn the relationship between the parameters and the constraint conditions. As a result, the machine learning model that has been pre-learned with less learning data can be further learned using the pre-data, thereby improving the accuracy of the prediction of the machine learning model. Since the prediction accuracy of the machine learning model is improved, the parameters representing the operating conditions of the air conditioning and cooling equipment 2C can be efficiently searched.

[0279] The parameter search device of Embodiment 6 may also be a parameter search device 1C that searches for parameters that represent the operating conditions of the air conditioning cooling and heating device 2C. In this case, the parameter search device 1C uses external information detected by an external sensor provided in the air conditioning cooling and heating device 2C to predict the relationship between the parameter that represents the operating conditions of the air conditioning cooling and heating device 2C, the evaluation value, and the external information, or the relationship between the parameter, the constraint condition, and the external information.

[0280] As an external sensor, for example, a temperature sensor or an infrared camera installed in the same room as the indoor unit 21 of the air conditioning device 2C is cited. The external information includes, for example, at least one of the dry-bulb temperature, wet-bulb temperature, gas pipe length, pipe (liquid pipe) length, refrigerant amount, and heat load of the air conditioning device 2C.

[0281] The information indicating the outdoor temperature may be the temperature information around the air conditioning device 2C acquired by the operation result collecting unit 11A from the network. In addition, the external information may be the information indicating the configuration of the air conditioning device 2C such as the number or capacity of the indoor units 21 .

[0282] The heat load is information related to indoor heat sources, which is determined by the outdoor temperature, people present in the room, or other electronic devices that may become heat sources. For example, when the electronic device is a smartphone, the operation result collection unit 11A calculates the heat load in the room where the air conditioning and heating device 2C is installed, using the temperature information detected by the indoor thermometer, the temperature information of the smartphone, the infrared camera image, and the weather information around the room where the air conditioning and heating device 2C is installed.

[0283] In addition, the operation result collection unit 11A may also calculate the heat load by considering the number of people in the room. The number of people may be calculated using the location information of the smartphone carried by each person, an infrared camera image, the entry and exit records of the room where the air conditioning and heating device 2C is installed, or the schedule information of each person. In addition, the operation result collection unit 11A may also calculate the heat load using a physical model related to the heat load in the room where the air conditioning and heating device 2C is installed, or a machine learning model learned in a manner to predict the heat load.

[0284] Information related to the dry-bulb temperature, wet-bulb temperature, air pipe length, piping length or piping height difference of the air conditioning and heating equipment 2C can also be collected by the operation result collection unit 11A from the directory information of the air conditioning and heating equipment 2C provided on the network.

[0285] In addition, the operation result collection unit 11A may collect the values ​​of each data as the above-mentioned external information, or may collect the index values ​​calculated based on these data.

[0286] By using external information indicating the installation environment of the air conditioning and heating device 2C, the parameter search device 1C can search and set appropriate parameters according to each destination provided by the air conditioning and heating device 2C. Thus, the air conditioning and heating device 2C can be operated in a manner that maximizes or minimizes the COP, cooling and heating capacity, or power consumption according to each installation environment.

[0287] Fig.17 2C is a conceptual diagram showing a refrigeration cycle of an air conditioning and cooling device 2C. The air conditioning and cooling device 2C is, for example, an air conditioner, a ventilation device, a sanitary device, a refrigerator, or a water heater. Fig.17As shown, the air conditioning device 2A includes an indoor unit 21 and an outdoor unit 22. Pipes 25A and 25B are pipes that connect the indoor unit 21 and the outdoor unit 22. Depending on the installation destination of the air conditioning device 2C, the lengths of the pipes 25A and 25B that connect the indoor unit 21 and the outdoor unit 22 may be different. The lengths of the pipes 25A and 25B are factors that have a great influence on the operation of the air conditioning device 2C. Therefore, depending on the lengths of the pipes 25A and 25B, the parameters, evaluation values ​​or constraints that represent the optimal operating conditions in the operation of the air conditioning device 2C are also different.

[0288] Therefore, the parameter search device 1C can also use the lengths of the pipes 25A and 25B as external information to predict the relationship between the parameters representing the operating conditions of the air conditioning device 2C, the evaluation value and the external information, or the relationship between the parameters, the constraints and the external information.

[0289] In this way, the parameter search device 1C searches for parameters indicating the operating conditions of the air-conditioning and cooling device 2C in consideration of standards such as JRA or the stability of the refrigeration cycle, thereby being able to provide a high-quality air-conditioning and cooling device 2C.

[0290] As described above, in the parameter search device 1A of the sixth embodiment, the mechanical device is the air conditioning device 2C. The constraint condition is that at least one of the cooling and heating capacity of the air conditioning device 2C, the power consumption of the air conditioning device 2C, the energy efficiency indicating the cooling and heating capacity corresponding to the power consumption of the air conditioning device 2C, the suction heating degree of the compressor 224 of the air conditioning device 2C, the inlet subcooling degree of the electronic expansion valves 212 and 221 of the air conditioning device 2C, the temperature and pressure of the refrigerant controlled by the compressor 224, the flow rate, temperature and pressure of the water to be cooled and heated controlled by the air conditioning device 2C, and the current flowing in the electronic circuit of the electronic substrate of the air conditioning device 2C is within a fixed range.

[0291] Thus, the parameter search device 1A can efficiently search for parameters indicating the operating conditions of the air-conditioning and cooling equipment 2A.

[0292] In the parameter search device 1C of the sixth embodiment, the external information includes at least one of the dry-bulb temperature, wet-bulb temperature, gas pipe length, liquid pipe length, refrigerant amount, and heat load of the air conditioning cooling and heating device 2C. Thus, the parameter search device 1C can efficiently search for parameters representing the operating conditions of the air conditioning cooling and heating device 2A.

[0293] Furthermore, the embodiments may be combined or arbitrary components of the embodiments may be modified, or arbitrary components may be omitted in the embodiments.

[0294] Industrial Applicability

[0295] The parameter search device of the present disclosure can be used, for example, to search for operating conditions of air-conditioning and heating equipment.

[0296] Description of Reference Numerals

[0297] 1, 1A, 1B, 1C parameter search device, 2 mechanical device, 2A, 2B, 2C air conditioning cooling and heating equipment, 3 advance data collection device, 4 external sensor, 11, 11A, 11B operation result collection unit, 12, 12A, 12B evaluation value acquisition unit, 13, 13A, 13B, 13C parameter search unit, 21 indoor unit, 22 outdoor unit, 23 flow meter, 24 air conditioning controller, 25A, 25B piping, 31 advance data acquisition unit, 32 advance data storage unit, 100 input interface, 101 output interface, 102 processing circuit, 103 processor, 104 memory, 111, 111A operation result acquisition unit, 112, 112A operation result storage unit, 113 sensor data acquisition unit, 114 sensor data storage unit, 121, 121A, 121B evaluation value calculation unit, 122, 122A, 122B evaluation value storage unit storage unit, 123, 123A constraint condition calculation unit, 131, 131A, 131B, 131C machine learning unit, 132, 132A, 132B, 132C search end determination unit, 133, 133A, 133B, 133C parameter determination unit, 211 indoor heat exchanger, 212, 221 electronic expansion valve, 222 refrigerant heat exchanger, 223 four-way valve, 224 compressor, 225 liquid storage device, 226 outdoor heat exchanger, 227A, 227B solenoid valves, 1311, 1311A, 1311B, 1311C learning prediction unit, 1312, 1312A, 1312B, 1312C prediction result storage unit, 1313 prior learning unit, 1331, 1331A, 1331B, 1331C search parameter calculation unit, 1332, 1332A, 1332B, 1332C parameter instruction unit.

Claims

1. A parameter search device, characterized in that: The parameter search device comprises: an operation result collecting unit that collects operation results including parameters indicating operation conditions of the mechanical device; an evaluation value acquisition unit that acquires an evaluation value of the parameter obtained using the operation result; as well as a parameter search unit that searches for the target parameter from among the parameters representing the operating conditions of the mechanical device, The parameter search unit predicts the relationship between the parameter in the search candidate and the evaluation value using a machine learning model, and when a plurality of search candidates regarding the parameter and the evaluation value are input, the machine learning model outputs the relationship between the parameter in the search candidate and the evaluation value, Comparing the prediction result of the machine learning model with the best search with the largest evaluation value among the searched parameters, and taking the number of search candidates whose evaluation values ​​exceed the best search as a search end indicator, Based on the comparison result between the search end index and the threshold, it is determined whether to continue searching for the parameter.

2. The parameter search device according to claim 1, characterized in that: The parameter search unit trains the machine learning model.

3. The parameter search device according to claim 1 or 2, characterized in that: The operation result collecting unit collects the operation result further including the constraint conditions to be satisfied in the operation of the mechanical device. The parameter search unit uses the following machine learning model to predict the relationship between the parameters, the evaluation values ​​and the constraints in the search candidates. When the machine learning model inputs multiple search candidates regarding the parameters, the constraints and the evaluation values, it outputs a predicted value representing the relationship between the parameters, the evaluation values ​​and the constraints in the search candidates.

4. The parameter search device according to claim 2, characterized in that: The parameter search unit obtains a pre-learned model that has previously learned the relationship between the parameter and the evaluation value, and trains the machine learning model using the pre-learned model.

5. The parameter search device according to claim 3, characterized in that: The parameter search unit obtains a pre-learned model that has previously learned the relationship between the parameters, the constraints, and the evaluation values, and trains the machine learning model using the pre-learned model.

6. The parameter search device according to any one of claims 1 to 5, characterized in that: The operation result collection unit further acquires external information including at least one of peripheral information of the mechanical device and information related to the operation task. The evaluation value acquisition unit calculates the evaluation value of the parameter using the external information.

7. The parameter search device according to claim 5, characterized in that: The operation result collection unit further acquires external information including at least one of peripheral information of the mechanical device and information related to the operation task. The evaluation value acquisition unit calculates the evaluation value of the constraint condition using the external information.

8. The parameter search device according to any one of claims 1 to 7, characterized in that: The mechanical device is an air conditioning heating and cooling device. The parameters include at least one of the opening of the electronic expansion valve, the opening of the solenoid valve, the operating frequency of the compressor, the air volume of the fan, the speed of the fan, the angle of the blades that determine the blowing direction, and the flow rate of the water to be cooled or heated.

9. The parameter search device according to claim 8, characterized in that: The evaluation value includes at least one of the cooling and heating capacity of the air-conditioning and cooling equipment, the energy efficiency of the cooling and heating capacity corresponding to the power consumption of the air-conditioning and cooling equipment, the comfort index of the air-conditioning and cooling equipment, the temperature of the air outlet of the air-conditioning and cooling equipment, the outlet temperature of the water subject to cooling and heating control of the air-conditioning and cooling equipment, and the concentration of carbon dioxide discharged through the cooling and heating control of the air-conditioning and cooling equipment.

10. The parameter search device according to claim 3, characterized in that: The mechanical device is an air conditioning heating and cooling device. The constraints are that at least one of the values ​​of the cooling and heating capacity of the air-conditioning and cooling equipment, the power consumption of the air-conditioning and cooling equipment, the energy efficiency representing the cooling and heating capacity corresponding to the power consumption of the air-conditioning and cooling equipment, the suction heating degree of the compressor of the air-conditioning and cooling equipment, the supercooling degree of the inlet of the electronic expansion valve of the air-conditioning and cooling equipment, the temperature and pressure of the refrigerant controlled by the compressor, the flow rate, temperature and pressure of the water controlled by the air-conditioning and cooling equipment, and the current flowing through the electronic circuit of the electronic substrate of the air-conditioning and cooling equipment is within a fixed range.

11. The parameter search device according to claim 6 or 7, characterized in that: The mechanical device is an air conditioning heating and cooling device. The external information includes at least one of a dry-bulb temperature, a wet-bulb temperature, a gas pipe length, a liquid pipe length, a refrigerant amount, and a heat load of the air conditioning cooling and heating equipment.

12. A parameter search method based on a parameter search device, characterized in that: The parameter search method comprises the following steps: The operation result collecting unit collects the operation results including parameters indicating the operation conditions of the mechanical device; An evaluation value acquisition unit acquires an evaluation value of the parameter obtained using the operation result; as well as The parameter search unit searches for the target parameter from among the parameters indicating the operating conditions of the mechanical device. The parameter search unit predicts the relationship between the parameter in the search candidate and the evaluation value using a machine learning model, and when a plurality of search candidates regarding the parameter and the evaluation value are input, the machine learning model outputs the relationship between the parameter in the search candidate and the evaluation value, Comparing the prediction result of the machine learning model with the best search with the largest evaluation value among the searched parameters, and taking the number of search candidates whose evaluation values ​​exceed the best search as a search end indicator, Based on the comparison result between the search end index and the threshold, it is determined whether to continue searching for the parameter.

Images