How to generate a trained model

Abstract

To provide an information processing device, etc. capable of suitably detecting an abnormality of a cold system.SOLUTION: An information processing device 1 includes: an acquisition section acquiring a measured measurement value from a cold system performing heat exchange; and an estimation section estimating whether or not the cold system is in an abnormal state and a type of the abnormal state from the measurement value acquired by the acquisition section on the basis of a plurality of learned models in which the measurement value and a state of the cold system during the measurement of the measurement value have been learned, the learned models prepared in accordance with the type of the abnormal state occurring in the cold system.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention , studies This article describes how to generate a trained model. [Background technology]

[0002] There are technologies to detect abnormalities in cooling and heating systems that perform heat exchange, such as air conditioning systems. For example, Patent Document 1 discloses a status monitoring device that generates frame images by dividing detection information such as current values ​​measured from an air conditioner at predetermined time intervals, constructs a model that learns the features of the frame images using machine learning, and detects abnormalities in the air conditioner. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent No. 6298562 Summary of the Invention [Problem to be solved by the invention]

[0004] However, in the invention of Patent Document 1, all kinds of anomalies that occur in air conditioners are learned using the same model, which may reduce the accuracy of anomaly detection.

[0005] In one aspect, an abnormality in a cooling and heating system can be suitably detected. How to generate a trained model The purpose is to provide. [Means for solving the problem]

[0006] One aspect of How to generate a trained model teeth, and acquiring training data for a measurement value measured from a cooling / heating system, the training data being assigned with information indicating the state of the cooling / heating system at the time the measurement value was measured and, if the state of the cooling / heating system is an abnormal state, information indicating the type of the abnormal state, and generating a plurality of trained models different for each type of abnormal state based on the training data, the trained models being trained with the measurement value as an input and the state of the cooling / heating system as an output, the trained models being different for each type of abnormal state, the measurement value in another type of abnormal state being input to each of the trained models thus generated to estimate whether or not the abnormal state is an abnormal state, calculating an index value indicating a correlation between each of the trained models based on an estimation result estimated from the measurement value in the other type of abnormal state, classifying the plurality of types of abnormal states into one or more groups based on the calculated index value, learning the training data for each classified group, and regenerating the trained models corresponding to each group. It is characterized by: Effect of the Invention

[0007] In one aspect, an abnormality in a cooling / heating system can be suitably detected. [Brief description of the drawings]

[0008] [Figure 1] FIG. 1 is a schematic diagram illustrating a configuration example of an anomaly detection system. [Diagram 2] FIG. 2 is a block diagram showing an example of the configuration of a server. [Diagram 3] FIG. 13 is an explanatory diagram showing an example of a record layout of a measurement value DB. [Figure 4] FIG. 1 is an explanatory diagram of an estimation model. [Diagram 5] FIG. 11 is an explanatory diagram regarding a generation process of an estimation model. [Figure 6] FIG. 13 is an explanatory diagram showing an example of a notification screen of an estimation result. [Figure 7] 13 is a flowchart showing a procedure for generating an estimation model. [Figure 8] 13 is a flowchart showing a procedure for a state estimation process. [Figure 9] FIG. 11 is a scatter plot showing the estimation results using the estimation model. [Figure 10] 13 is a flowchart showing a procedure for generating an estimation model according to the second embodiment. [Figure 11] 11 is a flowchart showing a procedure of a state estimation process according to the second embodiment. [Figure 12] FIG. 11 is an explanatory diagram relating to a mutual evaluation process between models. [Figure 13] FIG. 11 is an explanatory diagram showing a mutual evaluation result. [Figure 14] 13 is a flowchart showing a procedure for generating an estimation model according to the third embodiment. [Figure 15] 13 is a flowchart showing the procedure of a state estimation process according to the third embodiment. [Figure 16] FIG. 13 is an explanatory diagram showing an overview of a fourth embodiment. [Figure 17] 13 is a flowchart showing the procedure of a state estimation process according to the fourth embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] The present invention will now be described in detail with reference to the drawings showing embodiments thereof. (Embodiment 1) 1 is a schematic diagram showing a configuration example of an anomaly detection system. In this embodiment, an anomaly detection system that detects an anomaly in an air conditioning system 3 will be described. The anomaly detection system includes an information processing device 1, a terminal 2, and the air conditioning system 3. The information processing device 1 and the terminal 2 are communicatively connected via a network N such as the Internet.

[0010] The air conditioning system 3 is, for example, an air conditioning system installed in a building, and is an air conditioning system equipped with a refrigerator 32 as a heat source device that exchanges heat with an air conditioner 31. The refrigerator 32 is a so-called chiller, and exchanges heat with the air conditioner 31 via a refrigerant (for example, water). Note that the refrigerator 32 may be an air-cooled chiller.

[0011] In this embodiment, the air conditioning system 3 is given as an example of a cold energy system, but the cold energy system to be estimated is not limited to the air conditioning system 3, and may be, for example, a system that cools industrial machinery, measuring equipment, etc. using a refrigerator 32. In other words, the cold energy system may be any system that performs heat exchange, and its use is not limited to air conditioning.

[0012] In addition, in this embodiment, the air conditioning system 3 is described as being an air conditioning system equipped with a chiller 32, but the air conditioning system 3 may be an air conditioning system having only an outdoor unit and an indoor unit, and a configuration including a heat source device such as a chiller 32 is not essential.

[0013] The information processing device 1 is an information processing device capable of various information processing and transmitting and receiving information, for example, a server computer, a personal computer, etc. In this embodiment, the information processing device 1 is a server computer, and will be referred to as the server 1 below. The server 1 estimates the state of the air conditioning system 3 based on a predetermined measurement value measured from the air conditioning system 3. Specifically, as described later, the server 1 performs machine learning to learn the teacher measurement value and the state of the refrigerator 32 at the time of measuring the measurement value for each type of abnormal state occurring in the refrigerator 32, and generates a plurality of estimation models 141 (see FIG. 4 ). Then, the server 1 uses the plurality of estimation models 141 to estimate whether the refrigerator 32 is in an abnormal state, and if it is in an abnormal state, the type of the abnormal state.

[0014] The terminal 2 is, for example, a terminal device of an inspection company that performs maintenance and inspection of the air conditioning system 3, or an end user of the air conditioning system 3 (for example, an owner of a building), and is, for example, a personal computer, a smartphone, a tablet terminal, or the like. In this embodiment, the server 1 acquires measurement values ​​of the air conditioning system 3 from the terminal 2 and performs estimation. For example, the server 1 periodically acquires data of a daily management report that describes measurement values ​​measured from the chiller 32 and the contents of maintenance and inspection work (for example, repair, replacement, etc. of the chiller 32) from the terminal 2 and accumulates the data in a database. The server 1 performs learning using the data of the daily management report accumulated in the database as teacher data, and generates an estimation model 141. In addition, when estimating the state of the air conditioning system 3, the server 1 estimates the state of the air conditioning system 3 based on the measurement values ​​of the air conditioning system 3 that are periodically acquired from the terminal 2.

[0015] In this embodiment, processing is performed based on data recorded manually in the form of a daily management report, but the configuration may also be such that measurement values ​​are automatically obtained from the air conditioning system 3, such as a BEMS (Building Energy Management System).

[0016] Furthermore, in this embodiment, the processing is performed on the server 1 on the cloud, but the estimation model 141 may be installed on a local terminal 2 and processing related to state estimation of the air conditioning system 3 may be performed.

[0017] 2 is a block diagram showing an example of the configuration of the server 1. The server 1 includes a control unit 11, a main storage unit 12, a communication unit 13, and an auxiliary storage unit . The control unit 11 has one or more arithmetic processing devices such as a CPU (Central Processing Unit), an MPU (Micro-Processing Unit), a GPU (Graphics Processing Unit), etc., and performs various information processing, control processing, etc. by reading and executing a program P stored in the auxiliary storage unit 14. The main storage unit 12 is a temporary storage area such as an SRAM (Static Random Access Memory), a DRAM (Dynamic Random Access Memory), a flash memory, etc., and temporarily stores data necessary for the control unit 11 to execute arithmetic processing. The communication unit 13 is a communication module for performing processing related to communication, and transmits and receives information to and from the outside.

[0018] The auxiliary storage unit 14 is a non-volatile storage area such as a large-capacity memory or a hard disk, and stores a program P and other data necessary for the control unit 11 to execute processing. The auxiliary storage unit 14 also stores an estimation model 141 and a measurement value DB 142. The estimation model 141 is a trained model generated by machine learning, for example, a model related to a neural network. The estimation model 141 is expected to be used as a program module that is part of artificial intelligence software. The measurement value DB 142 is a database that stores measurement values ​​and the like measured from the refrigerator 32.

[0019] The auxiliary storage unit 14 may be an external storage device connected to the server 1. The server 1 may be a multi-computer consisting of a plurality of computers, or may be a virtual machine virtually constructed by software.

[0020] In the present embodiment, the server 1 is not limited to the above configuration, and may include, for example, an input unit for accepting operation input, a display unit for displaying images, etc. The server 1 may also include a reading unit for reading a portable storage medium 1a such as a CD (Compact Disk)-ROM or a DVD (Digital Versatile Disc)-ROM, and may read and execute the program P from the portable storage medium 1a. Alternatively, the server 1 may read the program P from a semiconductor memory 1b.

[0021] FIG. 3 is an explanatory diagram showing an example of a record layout of the measurement value DB 142. The measurement value DB 142 includes a chiller ID column, a date and time column, a measurement value column, a status column, and an operation column. The chiller ID column stores a chiller ID that is an identifier of the chiller 32 of the air conditioning system 3. The date and time column, the measurement value column, the status column, and the operation column store, in association with the chiller ID, the date and time when the measurement value was measured from the chiller 32, the measurement value, the status of the chiller 32 at the date and time, and the contents of the maintenance and inspection work performed when an abnormality was detected. The measurement value column stores the measurement values ​​measured from the chiller 32, which are the current value supplied to the chiller 32, the pressure values ​​of the high pressure side and the low pressure side of the compressor, the outside air temperature, the pressure difference which is the difference between the pressure values ​​of the high pressure side and the low pressure side of the compressor, and the temperature difference at the inlet and outlet of the chiller 32. The status column stores the status of the chiller 32 (whether it is in an abnormal state or a normal state other than an abnormal state), and, if it is in an abnormal state, a label indicating the type of the abnormal state. In the work column, the details of the work performed by the worker, such as repairing or replacing the refrigerator 32, are stored.

[0022] Fig. 4 is an explanatory diagram regarding the estimation model 141. An overview of the estimation model 141 will be described with reference to Fig. 4. The estimation model 141 is a model that takes the measured values measured from the air conditioning system 3 as input and outputs the estimation result of the state of the air conditioning system 3. Specifically, as shown in FIG. 4, the estimation model 141 is a neural network, and includes an input layer that receives the input of the measured values, an intermediate layer that extracts feature amounts from the measured values, and an output layer that outputs the estimation result based on the feature amounts extracted by the intermediate layer. Each layer is composed of one or more neurons.

[0023] In the present embodiment, a plurality of types of measured values measured from the refrigerator 32 of the air conditioning system 3 are input to the estimation model 141, and the estimation result of the state of the refrigerator 32 is output. The measured values input to the estimation model 141 are not particularly limited, but in the present embodiment, the current value A supplied to the refrigerator 32, the pressure differences dP on the high-pressure side and the low-pressure side of the compressor, and the pressure value P on the low-pressure side L , the pressure value P on the high-pressure side H , the temperature difference dT between the inlet and outlet of the refrigerator 32, and six types of measured values of the outside air temperature T0 are input. Note that these measured values are examples, and the number of measured values may be 5 or less, or 7 or more. Each type of measured value is input to each neuron of the input layer as input values x1 to x6. When inputting the measured values to the input layer, the standardized values of the measured values are input.

[0024] The output layer of the estimation model 141 includes two neurons that output an output value y1 representing the probability that the refrigerator 32 is in an abnormal state and an output value y2 representing the probability that it is in a normal state. The server 1 inputs each measured value to the input layer, performs calculations in the intermediate layer, and obtains the output values y1 and y2 from the output layer. For example, when y1 < y2, the server 1 determines that it is in an abnormal state, and when y1 ≥ y2, it determines that it is in a normal state.

[0025] During learning, the server 1 learns using the measurement values ​​at each measurement time point stored in the measurement value DB 142 and the state of the refrigerator 32 at each measurement time point as teacher data, and generates an estimation model 141. Specifically, the server 1 reads the above six types of measurement values ​​from the measurement value DB 142, inputs them as input values ​​x1 to x6 to the input layer, and obtains output values ​​y1 and y2 from the output layer. The server 1 compares the correct values ​​c1 and c2 with the output values ​​y1 and y2, setting c1=0.0 and c2=1.0 when the state of the refrigerator 32 at the measurement time point is abnormal, and setting c1=1.0 and c2=0.0 when the state is normal. The server 1 updates the weight w between neurons by the backpropagation method so that the two are approximate to each other. The server 1 sequentially inputs the measurement values ​​at each measurement time point included in the teacher data to optimize the weight w, and generates the estimation model 141.

[0026] 4 has only one intermediate layer, deep learning may be performed with multiple intermediate layers. Also, the estimation model 141 is not limited to a neural network, and may be a model based on other learning algorithms such as a support vector machine (SVM), a random forest, or a decision tree.

[0027] In addition, in this embodiment, an abnormality is estimated from a measurement value at one measurement time point, but the state of the refrigerator 32 may be estimated from measurement values ​​at multiple time points by using estimation model 141 as a model capable of handling time-series data, such as an RNN (Recurrent Neural Network).

[0028] Fig. 5 is an explanatory diagram regarding the generation process of the estimation model 141. Fig. 5 conceptually illustrates the generation of a separate estimation model 141 for each type of abnormal state that occurs in the refrigerator 32. The details of the machine learning process for generating the estimation model 141 will be described with reference to Fig. 5.

[0029] The server 1 performs machine learning to learn training data including measurement values ​​of the refrigerator 32 at each of a plurality of measurement time points and labels representing the state of the refrigerator 32 at each measurement point, and generates an estimation model 141. Specifically, as shown in Fig. 5, the training data is data in which, for each measurement value at each of a plurality of measurement time points, a label is assigned indicating whether the refrigerator 32 is in an abnormal or normal state, and, if it is in an abnormal state, the type of abnormal state. If it is in a normal state, a label of NF (Not Failure) is assigned, and if it is in an abnormal state, a label of Fx (x = 1, 2, 3 ...) representing the type of abnormality is assigned.

[0030] In this embodiment, eleven types of labels F1 to F11 are prepared as types of abnormal conditions. Detailed explanation and illustration of each label will be omitted, but for example, label F1 indicates a malfunction of the blower fan, and label F2 indicates an abnormality in the hot water. The number and contents of the types of abnormal conditions are not particularly limited and can be set arbitrarily.

[0031] The server 1 reads out the measurement values ​​in an abnormal state and the measurement values ​​in a normal state from the measurement value DB 142, and performs learning by providing these to the estimation model 141. In this case, the server 1 generates the estimation models 141 separately for estimating each type of abnormal state, and therefore performs learning by providing only the measurement values ​​in one type of abnormal state and the measurement values ​​in a normal state to one estimation model 141, without providing the measurement values ​​in other types of abnormal states.

[0032] 5 illustrates the generation of estimation models 141 corresponding to the labels F1 and F2 from the data of the air conditioning system 3 indicated by "air conditioner A". The teacher data shown in FIG. 5 includes two labels, F1 and F2, but when generating the estimation model 141 for estimating the label F1, the server 1 excludes the measurement values ​​assigned with the label F2, and provides only the measurement values ​​assigned with the labels F1 and NF to the estimation model 141 for learning. When learning the estimation model 141 for estimating the F2 label, the server 1 excludes the measurement values ​​assigned with the F1 label, and provides only the measurement values ​​assigned with the labels F2 and NF for learning.

[0033] In this way, the server 1 provides data of only one type of abnormal state to one estimation model 141, and generates estimation models 141 for estimating the labels F1, F2, F3, .... When multiple types of labels Fx are included in the training data, there is a risk of the learning accuracy being reduced if this is learned using the same estimation model 141. Therefore, in this embodiment, a reduction in learning accuracy is prevented by generating separate estimation models 141 for each type of abnormal state.

[0034] As described above, the measurement value assigned with the label Fx and the measurement value assigned with the label NF are provided to the estimation model 141 to perform learning. In this case, the server 1 performs learning by alternately and randomly extracting measurement values ​​assigned with the label Fx representing an abnormal state and measurement values ​​assigned with the label NF representing a normal state one by one. For example, when learning the estimation model 141 for estimating the label F1, the server 1 first randomly extracts measurement values ​​assigned with the label F1 from the teacher data and learns them, and then randomly extracts measurement values ​​assigned with the label NF and learns them. The server 1 alternately extracts measurement values ​​assigned with the labels F1 and NF and provides them to the estimation model 141 to perform learning until the number of learning times reaches a predetermined number of times.

[0035] In this way, the server 1 extracts data and performs learning so that the number of times the abnormal state and the normal state are learned is the same. In general, the air conditioning system 3 is rarely in an abnormal state compared to a normal state, so the teacher data contains overwhelmingly more data in a normal state than data in an abnormal state. Therefore, by learning both states the same number of times and learning alternately, the estimation accuracy can be improved.

[0036] In the above, data on abnormal and normal states are alternately learned one by one, but for example, data on both states may be alternately learned two by two. In other words, the server 1 only needs to be able to learn data on both states so that the number of data is the same, and the learning order is not particularly limited.

[0037] As described above, the server 1 alternately performs learning for each type of abnormal state so that the number of learning times for the abnormal state and the normal state is the same, and generates estimation models 141, 141, 141... for estimating various types of abnormal states. When actually estimating the state of the air-conditioning system 3 to be estimated, the server 1 acquires the measured values of the air-conditioning system 3 from the terminal 2 and performs estimation using the generated (learned) estimation models 141, 141, 141...

[0038] That is, the server 1 inputs each measured value measured from the refrigerator 32 of the air-conditioning system 3 to be estimated into each estimation model 141 for estimating labels F1, F2, F3... and obtains output values y1, y2 representing the probability of being in an abnormal state or a normal state from each estimation model 141. Then, based on the output from each estimation model 141, the server 1 estimates whether it is in an abnormal state or not, and if it is in an abnormal state, the type of the abnormal state. For example, the server 1 determines whether y1 < y2 for each estimation model 141, and if y1 < y2, it outputs the estimation results of each estimation model 141 in parallel, assuming that the abnormal state corresponding to the label Fx of that estimation model 141 has occurred.

[0039] For example, the server 1 may adopt the result of the estimation model 141 with the highest output value y1, determine the label Fx corresponding to the estimation model 141 as the type of the abnormal state occurring in the refrigerator 32, and integrate the estimation results of the plurality of estimation models 141, 141, 141... Various changes can be considered for how to process the output from each estimation model 141.

[0040] The server 1 notifies the terminal 2 of the above estimation result. Specifically, when it is estimated that the refrigerator 32 is in a normal state, the server 1 notifies the terminal 2 that the refrigerator 32 is normal. On the other hand, when it is estimated that the refrigerator 32 is in an abnormal state, the server 1 notifies that an abnormality has occurred in the refrigerator 32 and the type of the estimated abnormal state.

[0041] Fig. 6 is an explanatory diagram showing an example of a notification screen of the estimation result. The terminal 2 receives the output from the server 1 and displays the screen shown in Fig. 6. For example, as shown in Fig. 6, the terminal 2 displays a probability value (y2) for each type of abnormal state, and displays the estimation result of whether or not the state is abnormal.

[0042] In this case, it is preferable that the server 1 not only notifies the user of the abnormality of the refrigerator 32, but also notifies the user of a countermeasure against the estimated type of abnormal state. The countermeasure is, for example, a method of repairing or replacing the refrigerator 32, but the content of the countermeasure is not particularly limited.

[0043] For example, the server 1 stores in the measurement value DB 142 the details of the maintenance and inspection work performed when an abnormality occurred in the chiller 32, in association with the type of abnormal condition that occurred in the chiller 32. The server 1 reads out the corresponding past work details from the measurement value DB 142 according to the type of abnormal condition estimated above, and notifies the terminal 2. This enables the inspection company, end user, etc. of the air conditioning system 3 to quickly respond to a breakdown or the like of the air conditioning system 3.

[0044] In addition, it is preferable that the server 1 accepts an input for correcting the estimation result based on the estimation model 141 on the notification screen of Fig. 6, whether or not the estimation result based on the estimation model 141 is correct. For example, as shown in Fig. 6, the terminal 2 displays a tab 61 for correcting the estimation result, and accepts an operation input to the tab 61. This allows the server 1 to correct the estimation result (estimation model 141). For example, the server 1 stores the corrected estimation result as a label in the measurement value DB 142 in association with the measurement value, and uses the corrected estimation result as training data for re-learning to re-learn the estimation model 141. This allows the estimation accuracy based on the estimation model 141 to be improved.

[0045] As described above, according to the first embodiment, by using a plurality of different estimation models 141 for each type of abnormal state that occurs in the air conditioning system 3, an abnormality in the air conditioning system 3 can be suitably detected.

[0046] 7 is a flowchart showing the procedure of the generation process of the estimation model 141. The details of the machine learning process for generating the estimation model 141 will be described with reference to FIG. The control unit 11 of the server 1 acquires teacher data in which, for the measurement values ​​measured from the air conditioning system 3, information indicating the state of the air conditioning system 3 at the time of measurement and, if the state of the air conditioning system 3 is abnormal, information indicating the type of the abnormal state is assigned (step S11). The measurement values ​​are measurement values ​​related to the chiller 32 of the air conditioning system 3, and include the current value supplied to the chiller 32, the pressure values ​​and pressure difference of the high pressure side and low pressure side of the compressor, the temperature difference between the inlet and outlet of the chiller 32, the outside air temperature, etc. The control unit 11 acquires teacher data in which, for the measurement values, a label indicating whether the air conditioning system 3 was in an abnormal state or normal state at the time of measurement and, if it is in an abnormal state, a label indicating the type of the abnormal state is assigned.

[0047] The control unit 11 selects one type of abnormal state to be learned (step S12). The control unit 11 randomly extracts measurement values ​​labeled with the selected type of abnormal state from the teacher data (step S13). The control unit 11 provides the extracted measurement values ​​to the estimation model 141 and performs learning (step S14). Specifically, the control unit 11 inputs the measurement values ​​to the estimation model 141 to estimate the state of the air conditioning system 3, compares the estimation result with the correct label, and updates the weights of the estimation model 141.

[0048] Next, the control unit 11 randomly extracts measurement values ​​labeled as normal from the teacher data (step S15). The control unit 11 provides the extracted measurement values ​​to the estimation model 141 and performs learning (step S16).

[0049] The control unit 11 determines whether the measurement values ​​labeled with the type of abnormal state selected in step S12 have been learned a predetermined number of times (step S17). If it is determined that the measurement values ​​have not been learned the predetermined number of times (S17: NO), the control unit 11 returns the process to step S13. In this case, the control unit 11 extracts measurement values ​​at different measurement times from the teacher data (step S13) and performs learning (step S14). The control unit 11 repeats the process to generate an estimation model 141 corresponding to the type of abnormal state selected in step S12.

[0050] When it is determined that the measurement values ​​have been learned a predetermined number of times (S17: YES), the control unit 11 determines whether learning has been completed for all types of abnormal conditions (step S18). When it is determined that learning has not been completed (S18: NO), the control unit 11 returns the process to step S12. In this case, the control unit 11 selects a different type of abnormal condition (step S12) and alternately learns the measurement values ​​of abnormal conditions and normal conditions to generate the estimation model 141. When it is determined that learning has been completed for all types of abnormal conditions (S18: YES), the control unit 11 ends the series of processes.

[0051] 8 is a flowchart showing the procedure of the state estimation process. The process of estimating the state of the air conditioning system 3 using the estimation model 141 will be described with reference to FIG. The control unit 11 of the server 1 acquires the measurement values ​​of the air conditioning system 3 to be estimated (step S31). The control unit 11 inputs the acquired measurement values ​​to each of a plurality of estimation models 141 prepared according to the type of abnormal state, and estimates whether or not an abnormal state exists, and if an abnormal state exists, the type of the abnormal state (step S32).

[0052] The control unit 11 judges whether or not the air conditioning system 3 is in an abnormal state based on the estimation result in step S32 (step S33). If it is judged that the air conditioning system 3 is not in an abnormal state (S33: NO), the control unit 11 notifies the terminal 2 that the air conditioning system 3 is normal (step S34), and ends the series of processes.

[0053] If it is determined that there is an abnormal state (S33: YES), the control unit 11 notifies the terminal 2 that an abnormality has occurred in the air conditioning system 3 and the type of the estimated abnormal state (step S35). For example, the control unit 11 notifies that the estimated type of abnormality has occurred, and notifies the terminal 2 of a countermeasure for that type of abnormal state. The countermeasure is, for example, a method such as repairing or replacing the air conditioning system 3 (the refrigerator 32). For example, the control unit 11 refers to the measurement value DB 142 and notifies the terminal 2 of the contents of past maintenance and inspection work that was performed when the estimated type of abnormal state occurred. The control unit 11 ends the series of processes.

[0054] As described above, according to this embodiment 1, by using multiple trained models (estimation models 141) prepared according to the type of abnormal state, abnormality detection in the cooling system (air conditioning system 3) can be performed suitably.

[0055] Moreover, according to the first embodiment, it is also possible to present countermeasures against an estimated abnormality in the cooling system.

[0056] (Embodiment 2) In this embodiment, a mode will be described in which the type of measurement value to be input to the estimation model 141 is changed depending on the type of abnormal state to be estimated. Note that the same reference numerals will be used to designate the same contents as those in the first embodiment, and the description thereof will be omitted.

[0057] According to research by the inventors of the present application, it has been found that among the multiple input values ​​(measurement values) x1 to x6, some have a high correlation with the output values ​​y1 and y2 and others have a low correlation, depending on the estimation model 141 for estimating each type of abnormal state. Fig. 9 is a scatter diagram showing the estimation result by the estimation model 141. Fig. 9 shows a scatter diagram in the case where random numbers simulating measurement values ​​are input to the estimation model 141 for estimating the label F1 to estimate the label F1 or NF.

[0058] In each scatter plot in Figure 9, six types of measurements, A, dP, P L , P HThe results of the estimation of the F1 or NF labels are plotted with two arbitrary measurement values ​​of dP, P L , P H In the scatter plots with dT, T0 on the vertical and horizontal axes (second and subsequent rows in Fig. 9), the labels F1 and NF are scattered. On the other hand, in the scatter plot with the current value A on the horizontal axis (first row in Fig. 9), a classification boundary appears with a similar value as a boundary, and it can be seen that the current value A has a high correlation with the label F1.

[0059] In this way, there may be measurement values ​​that are highly correlated with the estimation result and measurement values ​​that are low in correlation, depending on the estimation model 141. Therefore, the estimation accuracy may be improved by inputting only measurement values ​​that are highly correlated and excluding other measurement values.

[0060] For example, when generating estimation models 141, 141, 141... for estimating each type of abnormal state from training data, the server 1 performs variable selection (feature selection) to select types of measurement values ​​predicted to have a high correlation with each type of abnormal state, and learns the selected types of measurement values ​​to generate each estimation model 141.

[0061] The specific method of variable selection is not particularly limited, but may be, for example, a filter method using a chi-square test, etc. Alternatively, the server 1 may use a method such as a wrapper method or an embedded method in which learning of teacher data and variable selection are performed simultaneously in parallel.

[0062] For example, before learning the training data, the server 1 performs statistical processing on the measurement values ​​to which labels Fx representing various types of abnormal states are assigned in the training data, and selects the type of measurement value that has a high correlation with the label Fx. L , P H , dT, T0 and the label F1 are calculated, and the current value A is selected as the measurement value having a high correlation according to the calculated statistical amount.

[0063] Similarly, the server 1 selects types of measurement values ​​that are highly correlated with the labels F2, F3, F4, etc. The server 1 generates an estimation model 141 for estimating the label Fx from the selected type of measurement value and the labels Fx and NF assigned to the measurement value. In the above example, the server 1 generates the estimation model 141 that receives the current value A as input and outputs the output values ​​y1 and y2 that indicate the probability of the label NF or F1.

[0064] In this way, the measurement values ​​used for estimation may be selected according to the type of abnormal state estimated by each estimation model 141. This is expected to improve the estimation accuracy, and even if some measurement values ​​are missing, for example, the estimation model 141 can perform estimation without any problems, thereby preventing a situation in which estimation of the state of the air conditioning system 3 cannot be performed at all.

[0065] 10 is a flowchart showing a procedure of the process of generating the estimation model 141 according to the embodiment 2. After acquiring the training data (step S11), the server 1 executes the following process. The control unit 11 of the server 1 selects, for each type of abnormal state, the type of measurement value to be input to the estimation model 141 for estimating each type of abnormal state (step S201). For example, the control unit 11 performs variable selection using a filter method or the like as described above, and selects the type of measurement value to be input to the estimation model 141. The control unit 11 transitions the process to step S12, and generates a plurality of estimation models 141, 141, 141... each having different types of measurement value to be input, depending on the type of abnormal state to be estimated.

[0066] 11 is a flowchart showing the procedure of the state estimation process according to embodiment 2. After acquiring the measured values ​​of the air conditioning system 3 that is the estimation target (step S31), the server 1 executes the following process. The control unit 11 of the server 1 selects different types of measurement values ​​according to the estimation models 141, 141, 141 ... for estimating each type of abnormal state (step S221). The control unit 11 inputs the selected types of measurement values ​​to the corresponding estimation models 141, and estimates the state of the air conditioning system 3 from the different types of measurement values ​​for each estimation model 141 (step S222). The control unit 11 transitions the process to step S33.

[0067] As described above, according to the second embodiment, by changing the type of measurement value to be input in accordance with each estimation model 141, it is possible to more suitably estimate each type of abnormal state.

[0068] (Embodiment 3) In the first embodiment, a configuration has been described in which the estimation model 141 is prepared (generated) for each type of abnormal condition. In the present embodiment, a configuration will be described in which the estimation models 141 generated for each type of abnormal condition are mutually evaluated, and a plurality of types of abnormal conditions are grouped to generate a more generalized estimation model 141.

[0069] Fig. 12 is an explanatory diagram of the mutual evaluation process between models. Fig. 12 illustrates a state in which a teacher measurement value to which a label F2 is assigned is input to an estimation model 141 for estimating a label F1, and the label F2 is estimated. An outline of this embodiment will be described with reference to Fig. 12.

[0070] In the first embodiment, the estimation model 141 is generated for each of the labels F1, F2, F3, .... On the other hand, even if the types of abnormal conditions are different, if the malfunctions (abnormalities) are similar, it is expected that the estimation model 141 can be generalized and the abnormal conditions can be estimated using the same estimation model 141. Therefore, the server 1 evaluates the correlation between the estimation models 141 generated for each label Fx, and generates the estimation model 141 that collectively estimates multiple types of abnormal conditions.

[0071] Specifically, the server 1 inputs a teacher measurement value to which a different type of label Fy (x≠y) is assigned to an estimation model 141 for estimating a label Fx. The server 1 then acquires output values ​​y1 and y2 relating to the state of the air conditioning system 3 from the output layer. The example in FIG. 12 illustrates how the server 1 inputs a measurement value to which a label F2 is assigned to an estimation model 141 for estimating a label F1, and acquires the output values ​​y1 and y2. The server 1 similarly inputs measurement values ​​for labels F3, F4, F5, ... to perform estimation.

[0072] In this way, the server 1 inputs the measurement values ​​of different labels Fy to the estimation model 141 for estimating the label Fx to obtain output values. Then, the server 1 calculates a predetermined index value representing the correlation between the estimation models 141 based on the output values ​​output from each estimation model 141. The index value is, for example, the recall rate between the estimation models 141 and the correlation coefficient of the output values. The recall rate is the rate at which the measurement value to which the label Fy is assigned is correctly estimated to be an abnormal state, and is calculated as recall rate=TP / (TP+FN), where TP is the number of times that the measurement value is correctly estimated to be an abnormal state and FN is the number of times that data that should be estimated to be an abnormal state is erroneously estimated to be a normal state. The correlation coefficient is, for example, the Pearson correlation coefficient, and is calculated from the covariance and standard deviation of the output values.

[0073] Fig. 13 is an explanatory diagram showing the mutual evaluation results. Fig. 13A shows a table showing the recall rates between estimation models 141, and Fig. 13B shows a table showing the correlation coefficients of output values. For ease of explanation, values ​​in the table are hatched.

[0074] 13A and 13B, the recall and correlation coefficients of the labels F1 and F2 are high, and the recall and correlation coefficients of the labels F6 to F9 are also high. Looking at the same, the labels can be roughly divided into a group of F1 to F4 and F6 to F9, and a group of F5, F10, and F11.

[0075] The server 1 classifies multiple types of abnormal states corresponding to each estimation model 141 into one or more groups according to the recall and correlation coefficients shown in Figures 13A and 13B. For example, the server 1 classifies abnormal states in which the recall and / or correlation coefficients between the estimation models 141 are equal to or greater than a certain value into the same group. Alternatively, the server 1 may display the recall and correlation coefficients shown in Figures 13A and 13B and accept manual input of grouping settings.

[0076] For example, the server 1 performs re-learning for each group classified as above, and regenerates the estimation model 141 that combines the multiple estimation models 141, 141, 141 .... For example, when the labels F1 to F4 and F6 to F9 are classified into the same group, the server 1 randomly extracts measurement values ​​to which each label Fx belonging to the group is assigned and measurement values ​​to which the label NF in a normal state is assigned, performs learning, and generates the estimation model 141 for estimating the labels F1 to F4 and F6 to F9.

[0077] In this case, for example, the server 1 may regard the labels F1 to F4 and F6 to F9 as one label (one type of abnormal state) and generate an estimation model 141 that estimates whether or not the label group consisting of the labels F1 to F4 and F6 to F9 is included. Alternatively, the server 1 may provide nine neurons in the output layer for calculating the probabilities of the labels F1 to F4, F6 to F9, and NF, and generate an estimation model 141 that can individually estimate the probability of the label being included. In this way, the regenerated estimation model 141 may be a model that estimates whether or not the corresponding abnormal state group (type group) is included, or may be a model that estimates whether or not the corresponding abnormal state type is included.

[0078] When actually estimating the state of the air conditioning system 3, the server 1 inputs the measurement values ​​to each estimation model 141 generated for each group, and outputs the estimation result for each group. In this way, it is also possible to group multiple types of abnormal states and construct a more generalized estimation model 141.

[0079] 14 is a flowchart showing a procedure of the generation process of the estimation model 141 according to the embodiment 3. After determining YES in step S18, the server 1 executes the following process. The control unit 11 of the server 1 calculates an index value representing the correlation with other estimation models 141 for each of the estimation models 141, 141, 141 ... generated according to the type of abnormal state (step S301). Specifically, as described above, the control unit 11 inputs the measurement value of the label Fy (x ≠ y) to the estimation model 141 for estimating the label Fx, and calculates the recall and correlation coefficient between the estimation models 141 from the output value.

[0080] The control unit 11 classifies the multiple types of abnormal conditions into one or multiple groups based on the calculated index values ​​(step S302). Then, the control unit 11 regenerates the estimation model 141 for each of the classified groups (step S303), and ends the series of processes.

[0081] 15 is a flowchart showing the procedure of the state estimation process according to embodiment 3. After acquiring the measured values ​​of the air conditioning system 3 that is the estimation target (step S31), the server 1 executes the following process. The control unit 11 of the server 1 inputs the acquired measurement values ​​to each estimation model 141, and estimates whether the air conditioning system 3 is in an abnormal state, and if so, the type of the abnormal state (step S321). Here, the control unit 11 performs estimation for each group classified in step S302. That is, the control unit 11 inputs the measurement values ​​to each estimation model 141 regenerated for each group as described above, and performs estimation for each group. The control unit 11 transitions the process to step S33.

[0082] As described above, according to the third embodiment, it is possible to integrate a plurality of types of abnormal states and construct a more generalized estimation model 141. This is expected to make it possible to apply the estimation model 141 to an air conditioning system 3 that has just started operation, for example, in which no measurement values ​​or failure history serving as teacher data have been accumulated.

[0083] (Embodiment 4) In the first embodiment, a description has been given of a mode for estimating a state in which an abnormality has actually occurred in the air conditioning system 3. In the present embodiment, a description will be given of a mode for estimating a state in which a sign of an abnormality occurring in the air conditioning system 3 is appearing.

[0084] Fig. 16 is an explanatory diagram showing an overview of the fourth embodiment. The overview of this embodiment will be described with reference to Fig. 16. In this embodiment, the server 1 estimates not only whether an abnormality is currently occurring in the air conditioning system 3 but also whether there are signs of an abnormality occurring, based on the output from the estimation model 141. For example, the server 1 learns training data to which a label Fxs indicating a symptom state before the occurrence of an abnormality is assigned in addition to NF indicating a normal state and a label Fx indicating that an abnormality is currently occurring, for training measurement values, and generates an estimation model 141 that outputs output values ​​y1, y2, y3 corresponding to each label.

[0085] The server 1 inputs the measured values ​​into the estimation model 141 generated above, and estimates the state of the air conditioning system 3. For example, the server 1 adopts the label with the largest value among the labels Fx, Fxs, and NF as the estimation result. That is, when the label Fx is the highest, it is estimated that an abnormality is actually occurring, and when the label Fxs is the highest, it is estimated that the state is a symptom of an abnormality. The server 1 outputs the estimation result to the terminal 2, and when the state is a symptom of an abnormality, it presents a message that there is a symptom of an abnormality.

[0086] In addition, when it is estimated that the symptom state is present, it is preferable that the server 1 presents not only the symptom state but also the time when the abnormality is predicted to occur. For example, the server 1 stores the period between the time when the label Fxs representing the symptom state first appears in the training data and the time when Fx representing the state in which the abnormality is actually occurring first appears, and outputs the period to the terminal 2 as a reference value until the abnormality occurs. Alternatively, the server 1 may make the estimation model 141 learn the period from the time of the symptom state to the time when the abnormality occurs. This allows the inspection company or the end user to know the time to take measures, improving convenience.

[0087] In the above, a label representing a symptom state is added to the teacher data to estimate a symptom of an abnormality, but the present embodiment is not limited to this. For example, the server 1 may use the estimation model 141 having only output values ​​y1 and y2 as in the first embodiment, and estimate whether or not the symptom stage is present depending on the magnitude of the output value y2 representing the probability of an abnormal state. For example, the server 1 estimates that an abnormality is actually occurring when y2>0.9 (90%), and estimates that the symptom state is present when y2≦0.9. In this way, the configuration of learning teacher data to which a label representing a symptom state is added is not essential, and whether or not the symptom state is present may be estimated based on the magnitude of the output value of the estimation model 141.

[0088] 17 is a flowchart showing the procedure of the state estimation process according to embodiment 4. After acquiring the measured values ​​of the air conditioning system 3 that is the estimation target (step S31), the server 1 executes the following process. The control unit 11 of the server 1 inputs the acquired measurement values ​​to each estimation model 141 and estimates whether an abnormality is currently occurring or whether the state is a symptom of an abnormality occurring, together with the type of abnormal state (step S401). The control unit 11 shifts the process to step S33.

[0089] When it is determined that the air conditioning system 3 is not currently in an abnormal state (S33: NO), the control unit 11 determines whether or not the state is a symptom of an abnormality occurring, based on the estimation result in step S401 (step S402). When it is determined that the state is not a symptom state (S402: NO), the control unit 11 shifts the process to step S34. When it is determined that the state is a symptom state (S402: YES), the control unit 11 notifies the terminal 2 that there is a symptom of an abnormality (step S403). For example, the control unit 11 notifies the terminal 2 of the type of abnormality that is estimated to be a symptom and the time when the abnormality is predicted to occur. The control unit 11 ends the series of processes.

[0090] As described above, according to the fourth embodiment, the occurrence of an abnormality in the air conditioning system 3 can also be predicted.

[0091] The embodiments disclosed herein are illustrative in all respects and should not be considered as limiting. The scope of the present invention is defined by the claims, not by the above meaning, and is intended to include all modifications within the scope and meaning equivalent to the claims. [Explanation of symbols]

[0092] 1. Server (information processing device) 11 Control section 12 Main memory 13. Communications Department 14 Auxiliary storage P Program 2. Terminal 3. Air Conditioning System 31 Air conditioner 32 Refrigeration Machine

Claims

1. acquiring teacher data in which, for a measurement value measured from a refrigeration system, information is added that indicates the state of the refrigeration system at the time the measurement value was measured, and, if the state of the refrigeration system is an abnormal state, information that indicates the type of the abnormal state; generating a plurality of trained models based on the teacher data, the trained models having the measurement values ​​as input and the state of the refrigeration system as output, the trained models being different for each type of the abnormal state; The measurement values ​​in the other types of abnormal states are input to each of the generated trained models to estimate whether or not the other types of abnormal states are in an abnormal state; Calculating an index value representing a correlation between each of the trained models based on an estimation result estimated from the measurement value in another type of abnormal state; classifying the plurality of types of abnormal conditions into one or a plurality of groups based on the calculated index values; The training data is learned for each classified group, and the trained model corresponding to each group is regenerated. A method for generating a trained model, comprising the steps of: having a computer execute processing.

2. The trained models are generated by extracting an equal number of the measurement values ​​in the abnormal state of the type corresponding to each trained model and the measurement values ​​in a normal state other than the abnormal state from the training data. The method for generating a trained model according to claim 1 .

Images