Learning Device, Monitoring Device, and Air Conditioning System

The present disclosure relates to a learning device, a monitoring device, and an air conditioning system.

Conventionally, a method of detecting leakage of refrigerant in a cooling system based on a value serving as an indicator of an amount of refrigerant introduced therein has been reviewed, and Japanese Patent No. 6791429 discloses an exemplary refrigerant amount determination device configured to facilitate determination of an amount of refrigerant.

The refrigerant amount determination device includes: a calculation unit configured to calculate a refrigerant amount indicator value from operation data of an air conditioning system; an inference unit configured to infer information about correction of the refrigerant amount indicator value using a correction model and at least one of the operation data and the calculated refrigerant amount indicator value; and a determination unit configured to determine the amount of the refrigerant of the air conditioning system based on the information about correction of the refrigerant amount indicator value.

In the refrigerant amount determination device disclosed in Japanese Patent No. 6791429, an influence of a decrease in the amount of the refrigerant over an air conditioning performance cannot be precisely known and a timing for maintenance cannot be known, disadvantageously. For example, when an amount of leakage of the refrigerant is small to be 10%, the air conditioning performance is not decreased so much even though it is known from the refrigerant amount indicator that the refrigerant is being leaked, with the result that it cannot be determined whether the location of the leakage should be checked urgently or the current situation should be kept and observed for a while.

The present disclosure has been made to solve the above-described problem, and has an object to provide a learning device so as to obtain an inference model useful to improve accuracy of detecting an abnormality of an air conditioning system.

The present disclosure relates to a learning device configured to learn a condition of an air conditioning apparatus in which refrigerant circulates. Operation data of the air conditioning apparatus includes a first data group and a second data group that does not contain a same data element as a data element of the first data group. The learning device includes: a first calculation unit configured to calculate a first feature amount from the first data group of the air conditioning apparatus during a learning period; and a learning unit configured to generate a first inference model that infers a first normal range of the first feature amount from the second data group by performing supervised learning using the second data group with the first feature amount being set as truth data.

According to the learning device of the present disclosure, since the model that infers the normal range of the feature amount from the behavior of the data other than the data that can be directly used for the calculation of the feature amount among the operation data is generated, possibility of detecting an abnormality accurately at an early stage can be increased.

Hereinafter, embodiments of the present disclosure will be described in detail with reference to figures. It should be noted that the same or corresponding portions in the figures are denoted by the same reference characters and will not be described repeatedly in principle.

is a block diagram showing a configuration of an air conditioning system according to a first embodiment. An air conditioning systemincludes: an abnormality detection system; and an air conditioning apparatushaving a condition monitored by abnormality detection system. As shown in, abnormality detection systemis connected to air conditioning apparatusvia a network.

Abnormality detection systemincludes a CPU (Central Processing Unit), a memory(a ROM (Read Only Memory) and a RAM (Random Access Memory)), an input/output buffer (not shown), and the like. CPUloads a program stored in the ROM into the RAM or the like and executes the program. The program stored in the ROM is a program in which a processing procedure of abnormality detection systemis written. Abnormality detection systemmonitors each device in air conditioning apparatusin accordance with such a program. This control is not limited to a process by software, and the process can also be performed by dedicated hardware (electronic circuit). It should be noted that abnormality detection systemmay be constructed in a server connected to network.

Air conditioning apparatusincludes a plurality of indoor units, an outdoor unit, and a controller. Each of the plurality of indoor unitsis disposed in an indoor space and is connected to outdoor unitby a liquid pipe and a gas pipe through each of which refrigerant passes. Outdoor unitis disposed in a space (outdoor space) outside the indoor space. It should be noted that the number of indoor unitsincluded in air conditioning apparatusmay be one or two.

Outdoor unitincludes a compressor, an outdoor heat exchanger, and an outdoor fan. Each of the plurality of indoor unitsincludes an expansion valve and an indoor heat exchanger. The refrigerant is supplied from the compressor included in outdoor unitto each of the plurality of indoor units. The refrigerant circulates between each of the plurality of indoor unitsand outdoor unit.

Controllerincludes a thermostat and integrally controls air conditioning apparatus. Controlleris connected to an abnormality detection systemvia network. Networkincludes the Internet and a cloud system. It should be noted that networkmay be a LAN (local area network).

Controllerincludes a CPU, a memory(a ROM and a RAM), an input/output buffer (not shown), and the like. CPUloads a program stored in the ROM into the RAM or the like and executes the program. The program stored in the ROM is a program in which a processing procedure of controlleris written. Controllerperforms control of each device in air conditioning apparatusin accordance with such a program. This control is not limited to a process by software, and the process can also be performed by dedicated hardware (electronic circuit).

is a functional block diagram showing a configuration of air conditioning apparatusof. As shown in, outdoor unitincludes a compressor, an outdoor heat exchanger, a four-way valve, an outdoor fan, an accumulator, temperature sensorsto, pressure sensors,, and a humidity sensor.

Each of the plurality of indoor unitsincludes an expansion valve, an indoor heat exchanger, an indoor fan, and temperature sensorsto. It should be noted that expansion valveincludes, for example, an LEV (linear expansion valve).

Operation modes of air conditioning apparatusinclude a heating mode, a cooling mode, and a defrosting mode. In the heating mode, four-way valveconnects a discharge port of compressorand indoor heat exchanger, and connects outdoor heat exchangerand a refrigerant inlet of accumulator. In the heating mode, the refrigerant circulates through compressor, four-way valve, indoor heat exchanger, expansion valve, and outdoor heat exchangerin this order. In each of the cooling mode and the defrosting mode, four-way valveconnects the discharge port of compressorand outdoor heat exchanger, and connects indoor heat exchangerand the refrigerant inlet of accumulator. In each of the cooling mode and the defrosting mode, the refrigerant circulates through compressor, four-way valve, outdoor heat exchanger, expansion valve, and indoor heat exchangerin this order.

Temperature sensormeasures a temperature (room temperature TH) of air suctioned into indoor heat exchanger, and outputs the temperature to controller. Temperature sensors,respectively measure temperatures (indoor liquid temperature THand indoor gas temperature TH) of the refrigerant before and after passing through indoor heat exchanger, and output the temperatures to controller.

Temperature sensormeasures a temperature (discharge temperature TH) of the refrigerant discharged from compressor, and outputs the discharge temperature to controller. Temperature sensormeasures a temperature (suction temperature TH) of the refrigerant suctioned into compressorvia accumulator, and outputs the suction temperature to controller. Temperature sensormeasures a temperature THof the liquid refrigerant in the pipe that connects outdoor heat exchangerand liquid pipe, and outputs the temperature to controller.

Pressure sensormeasures a pressure (discharge pressure HS) of the refrigerant discharged from compressor, and outputs discharge pressure HSto controller. Pressure sensormeasures a pressure (suction pressure LS) of the refrigerant suctioned into compressor, and outputs suction pressure LS to controller.

Hereinafter, control during cooling will be described as a representative example. Controllercontrols an amount of the refrigerant discharged by compressorper unit time by controlling an operation frequency fCOMP of compressorso as to allow a suction saturation gas temperature to become a target temperature. Controllercontrols a degree of opening Li of expansion valveso as to allow a degree of superheat SH (=TH-TH) of the refrigerant at the outlet of indoor heat exchangerto become a target value. Controllerswitches a circulation direction of the refrigerant by controlling four-way valveto provide a flow path indicated by the solid lines. Controllercontrols an amount of air sent by outdoor fanper unit time by controlling a rotation frequency fFANo of outdoor fanso as to allow a discharge saturation gas temperature to become a target value. Controllercontrols a rotation frequency fFANi of indoor fanso as to attain an amount of air set by a user. Controllerassociates, with a time at the measurement, operation data reflecting the condition of the air conditioning system, and transmits it to abnormality detection system.

is a diagram showing exemplary operation data reflecting the condition of air conditioning apparatusof. As shown in, the operation data includes, for example, an outdoor air temperature, the discharge temperature (TH), the evaporation temperature (TH), a condensation temperature, the suction temperature (TH), a send-out temperature, the high pressure (HS), the low pressure (LS), the operation frequency (fCOMP) of compressor, the degree of opening of expansion valve, the operation modes, an operation state (operation, halt, or standby), and the rotation speed (fFANo, fFANi)) of each of outdoor fanand indoor fan, the temperature (setting temperature) of the indoor space as set by the user, a current value of an inverter of compressor, a voltage value of the inverter, a temperature of a heat sink included in outdoor unit, and the temperature (liquid pipe temperature TH) of the liquid pipe that connects outdoor unitand indoor unit(pipe through which the liquid refrigerant flows). It should be noted that the operation frequency of compressor, the degree of opening of expansion valve, and the rotation speed of outdoor fanare basic operation amounts in VRF (Variable Refrigerant Flow) control.

An environment in which air conditioning apparatusis operated may have characteristics specific to the environment (for example, a length of the refrigerant pipe, a type of indoor unit, the number of indoor units, and a difference in height between indoor unitand outdoor unit). Therefore, a determination criterion (for example, a threshold value) for detecting an abnormality of air conditioning apparatuscan be different for each environment in which air conditioning apparatusis operated. Therefore, when the same determination criterion is used regardless of the environment in which air conditioning apparatusis operated, accuracy of detecting an abnormality of air conditioning apparatusmay be decreased.

Therefore, in abnormality detection system, when a normal condition is confirmed during a trial operation after installation of the air conditioning system, the normal condition is regarded as being kept for a subsequent certain period. Then, a trained model is generated in which a relation is learned between operation data of air conditioning apparatusduring the certain period and a normal value of a condition indicator value (specific parameter) representing the condition of air conditioning apparatusand corresponding to the operation data. By using the trained model, an abnormality of air conditioning apparatuscan be detected in accordance with a determination criterion adapted to the environment in which air conditioning apparatusis operated. As a result, the accuracy of detecting an abnormality of the air conditioning system can be improved.

is a block diagram showing a configuration of abnormality detection systemof. In abnormality detection system, operation data is obtained in each of a plurality of consecutive operation periods (for example, one day, one week, or one month), and trained inference models M, Mare constructed using training data including the operation data. The plurality of consecutive operation periods may be the same period (predetermined period) or may be different periods.

As shown in, abnormality detection systemincludes a learning deviceand a monitoring device. Learning deviceincludes a calculation unitA and a learning unit. Monitoring deviceincludes a calculation unitB, an inference unit, a determination unit, and a display unit.

Calculation unitA and calculation unitB basically perform the same calculation and are only different from each other in that they are used respectively at the time of learning and at the time of monitoring. Therefore, calculation unitA may also serve as calculation unitB.

Calculation unitA calculates a feature amount FAfrom operation data DAat the time of the normal condition. Learning unitperforms machine learning with feature amount FAof an element device and operation data DAof the element device at the time of the normal condition so as to construct inference model M. Inference unitinfers a normal range (upper and lower limit values) FBof the feature amount by inputting, to trained inference model M, operation data DBat the time of determination. Further, calculation unitA calculates feature amount FAfrom operation data DAat the time of the normal condition. Learning unitperforms machine learning with feature amount FAof an element device and operation data DAof the element device at the time of the normal condition so as to construct inference model M. Inference unitinfers a normal range (upper and lower limit values) FBof the feature amount by inputting, to trained inference model M, operation data DBat the time of determination.

Calculation unitB calculates a feature amount FBfrom operation data DBat the time of determination. Calculation unitB calculates a feature amount FBfrom operation data DBat the time of determination. Determination unitincludes a data processing unitand a data processing unit. Data processing unitcounts the number of pieces of data (data falling out of the normal range) for which feature amount FB(degree of supercooling SC of the refrigerant at the outlet of the condenser) calculated by calculation unitfrom operation data DBat the time of determination exceeds or falls below normal range FBinferred by trained inference model Mfrom the operation data at the time of poor condition. Data processing unitcounts the number of pieces of data for which feature amount FB(heat exchange performance Qo of the heat exchanger) calculated from operation data DBat the time of determination is equal to or less than a predefined performance. Determination unitcalculates a ratio by dividing, by the determined number of pieces of operation data, the number of pieces of data for which feature amount FB(degree of supercooling SC) is outside the normal range (falls out of the normal range) and feature amount FB(heat exchange performance Qo) is decreased by a predefined amount (falls out of the normal range) at the same time. This ratio represents a ratio of times in a day at each of which an operation involving a heat exchange performance Qo decreased to be lower than that in the normal condition takes place due to a poor condition of the element device, and is referred to as “performance decrease operation ratio” in the present specification. Determination unitdetermines whether or not the performance decrease operation ratio is equal to or more than a threshold value.

Display unitdisplays trend data indicating a change of the amount of the refrigerant and a change of the performance decrease operation ratio per day for each element device (outdoor heat exchanger, indoor heat exchanger), and displays that maintenance is necessary when the performance decrease operation ratio is equal to or more than the threshold value.

CPUshown inis operated as calculation unitsA,B, learning unit, inference unit, and determination unitin accordance with corresponding programs. Further, memorystores an operation data set of air conditioning apparatusand inference models M, M.

Each of inference models M, Mis a regression model that includes a neural network and that infers a normal value of the condition indicator value of air conditioning apparatusfrom the operation data of air conditioning apparatus. Each of inference models M, Mmay be a classification model that infers a level (classification) of the condition indicator value. The normal value of the condition indicator value may be each of a maximum value and a minimum value of a confidence interval that the condition indicator value can have when air conditioning apparatusis normal. Further, a range having the normal value as a median value (for example, a range of ±10% of the normal value) may be employed as the confidence interval.

It should be noted that a general AI (Artificial Intelligence) technology can be applied to clustering and weighting of the parameters included in the operation data.

Learning unitperforms machine learning onto inference models M, Mso as to construct trained inference models M, M. A machine learning algorithm used by learning unitmay be a known algorithm such as supervised learning, semi-supervised learning, unsupervised learning, or reinforcement learning. Further, as the machine learning algorithm, deep learning in which extraction of the feature amount itself is learned can be used, or the machine learning may be performed in accordance with another known method such as a neural network, genetic programming, inductive logic programming, or a support vector machine. Hereinafter, a case where the supervised learning is applied to a neural network will be described.

Calculation unitA calculates a condition indicator value (SC) during a first operation period from operation data DAincluded in operation data DA. Calculation unitA calculates a condition indicator value (Qo) during the first operation period from operation data DAincluded in operation data DA. Learning unitperforms the supervised learning onto inference model Musing training data having the condition indicator value during the first operation period as truth data (teaching data) for the first operation data. Learning unitconstructs trained inference models M, M, and stores the models into memory. It should be noted that since the first operation period is a period immediately after air conditioning apparatusis confirmed to be operated in the trial operation after being installed in the installation location, there is a low possibility that air conditioning apparatushas an abnormality. Since it can be assumed that air conditioning apparatusin the first operation period is normal, the condition indicator value during the first operation period is set as the truth data for the first operation data. In the below-described example, the first operation period is 365 days immediately after the installation.

is a flowchart for illustrating a detail of the learning process performed by abnormality detection system. When the learning process is started, abnormality detection systemshown inobtains air conditioning operation data DA for 365 days immediately after the installation in step S. Air conditioning operation data DA includes a first data group DAand a second data group DAthat does not contain the same data element as a data element of first data group DA. Further, air conditioning operation data DA includes a third data group DAdifferent from first data group DAand a fourth data group DAthat does not contain the same data element as a data element of third data group DA.

Then, in a step S, calculation unitA calculates a feature amount FAfrom first data group DAand calculates feature amount FAfrom third data group DA. Examples of feature amount FAinclude degree of supercooling SC of the refrigerant at the outlet of the condenser. Examples of feature amount FAinclude heat exchange performance Qo of outdoor heat exchanger.

For example, degree of supercooling SC at the outlet of the condenser can be calculated by the following formula (1):

Here, Tc represents the discharge saturation gas temperature and is a value determined by discharge pressure HS. THrepresents a liquid refrigerant temperature.

Further, for example, heat exchange performance Qo can be calculated by the following formula (2):

Here, Gr represents a refrigerant circulation amount and is a value determined by operation frequency fCOMP of compressor, and Hd represents a specific enthalpy of the inlet portion of outdoor heat exchangerand is a value determined by pressure HSand temperature TH. Further, Hco represents a specific enthalpy of the outlet of outdoor heat exchangerand is a value determined by pressure HSand temperature TH.

The upper part of Table 1 below shows: operation data DA used when training inference model Mfor determining leakage of the refrigerant; operation data DBused as an input to trained inference model M; and operation data DBused by calculation unitB at the time of determination.

Further, the lower part of Table 1 shows: operation data DA used when training inference model Mfor determining insufficiency of the heat exchange performance of outdoor heat exchanger; operation data DBused as an input to trained inference model M; and operation data DBused by calculation unitB at the time of determination.

Then, in a step Sof, learning unitconstructs inference model (regression model) Mof the objective variable (feature amount FA) corresponding to the explanatory variable (operation data group DA). Similarly, learning unitconstructs inference model (regression model) Mof the objective variable (feature amount FA) corresponding to the explanatory variable (operation data group DA). On this occasion, learning unitconstructs each inference model with the number of feature amounts×3 (upper limit value, median value, and lower limit value). For example, the upper limit value may be 97.5% of the confidence interval and the lower limit value may be 2.5% of the confidence interval (corresponding to 2σ).

is a diagram showing an exemplary neural network Nwincluded in inference model Mconstructed in step Sof. As shown in, neural network Nwincludes an input layer X, an intermediate layer (hidden layer) Y, and an output layer Z. Input layer Xincludes neurons X, X, X. Intermediate layer Yincludes neurons Y, Y. Output layer Zincludes a neuron Z. Input layer Xand intermediate layer Yare fully connected to each other. Intermediate layer Yand output layer Zare fully connected to each other. Neural network Nwmay include two or more intermediate layers.

When a plurality of inputs are respectively input to neurons Xto Xof input layer X, the values are multiplied by weights w, w, w, w, w, wand are input to neurons Y, Yof intermediate layer Y. The outputs from neurons Y, Yare multiplied by weights w, wand are output from neuron Zof output layer Z. The output result from output layer Zdiffers depending on the values of weights wto w, w, w. The weights and biases of neural network Nware updated by backpropagation of an error between the truth data and the result output from the output layer as a result of inputting the operation data to the input layer such that the result becomes close to the truth data.