Learning device and abnormality determination system

Abstract

To minimize a load on each of a learning device which learns an abnormality determination model and an abnormality determination system which performs abnormality determination.SOLUTION: A learning device includes a learning data acquisition unit that acquires learning data including plural time-series operation data items of an object apparatus, a data preprocessing unit that performs preprocessing on operation data using a change point of the operation data, and a model generation unit that uses the preprocessed operation data to generate an abnormality determination model which employs as input data the plural time-series operation data items of the object apparatus and regards as output data a result of abnormality determination of the object apparatus.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present disclosure relates to a learning device that learns an abnormality determination model and an abnormality determination system that performs abnormality determination using the abnormality determination model. [Background technology]

[0002] It is known that, in a conventional refrigeration cycle device, the operating state of the refrigeration cycle device is monitored based on measurement data and control signal data from various sensors, and if there is a sign of an abnormality, the operation is stopped or control is performed to avoid a failure. Typically, an abnormal state in the refrigeration cycle device is determined by comparing the measurement data and control signal data from the various sensors with thresholds set for each abnormal state. In this case, it is necessary to set each abnormal state and a large number of thresholds for determining each abnormal state before shipping the refrigeration cycle device.

[0003] In addition to comparing the operating state with a threshold value indicating an abnormal state, it is also known to use a neural network to perform calculations using a nonlinear function on measurement data and control signal data to determine abnormalities. For example, Patent Document 1 proposes using a neural network to diagnose faults in a compressor of a refrigeration cycle device. A neural network, which allows a computer to function as a calculation unit that calculates an output for a predetermined input, is capable of processing information by learning many actual examples. Furthermore, because it can perform complex information processing at high speed, various applications have been attempted in fields such as image recognition, voice recognition, voice synthesis, and automatic translation. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 7049363 Summary of the Invention [Problem to be solved by the invention]

[0005] Typically, high-dimensional data consisting of sensor measurement data and control signal data is very long, so a large-capacity storage device is required to store this data for a long period of time. Furthermore, a high-performance computing device is required to input the large-capacity data into a neural network and perform calculations for anomaly detection. In other words, the device that trains the model for anomaly detection and the device that performs anomaly detection using the model must be large-capacity and high-performance, which increases costs.

[0006] The present disclosure is intended to solve the above-mentioned problems, and aims to reduce the load on a learning device that learns an abnormality determination model and an abnormality determination system that performs abnormality determination. [Means for solving the problem]

[0007] The learning device according to the present disclosure includes a learning data acquisition unit that acquires learning data including multiple pieces of time-series operating data of a target device, a data pre-processing unit that pre-processes the operating data using change points in the operating data, and a model generation unit that uses the pre-processed operating data to generate an abnormality determination model that uses the multiple pieces of time-series operating data of the target device as input data and the abnormality determination result of the target device as output data.

[0008] The abnormality determination system according to the present disclosure includes an operating data acquisition unit that acquires multiple pieces of time-series operating data of a target device, a data pre-processing unit that pre-processes the operating data using change points in the operating data, and an abnormality determination unit that determines an abnormality in the target device using a trained abnormality determination model that uses the pre-processed operating data as input data and an abnormality determination result of the target device as output data. [Effects of the Invention]

[0009] According to the present disclosure, by preprocessing the driving data using change points in the driving data, it is possible to compress the data for learning and abnormality determination, thereby reducing the load on the learning device and the abnormality determination system. [Brief explanation of the drawings]

[0010] [Figure 1] 1 is a control block diagram of a learning device according to a first embodiment. [Figure 2] 4 is a flowchart showing the flow of a learning process performed by the learning device according to the first embodiment. [Figure 3] FIG. 4 is a diagram illustrating a data flow in the first learning process according to the first embodiment. [Figure 4] FIG. 4 is a diagram showing an example of operation data of the refrigeration cycle device. [Figure 5] FIG. 5 is a diagram showing an abnormality level α of the operational data shown in FIG. 4. [Figure 6] 4 is an example of preprocessed data processed by a data preprocessing unit of the learning device according to the first embodiment. [Figure 7] FIG. 10 is a diagram illustrating a data flow in the second learning process according to the first embodiment. [Figure 8] 1 is a schematic configuration diagram of an abnormality determination system according to a first embodiment. [Figure 9] 1 is a refrigerant circuit diagram of a refrigeration cycle device according to a first embodiment. [Figure 10] 4 is a flowchart showing the flow of an abnormality determination process performed by the abnormality determination device according to the first embodiment. [Figure 11] FIG. 3 is a diagram illustrating a data flow in the abnormality determination process according to the first embodiment. [Figure 12] FIG. 10 is a schematic configuration diagram of an abnormality determination system according to a second embodiment. [Figure 13] 10 is a flowchart showing the flow of an abnormality determination process performed by the abnormality determination device according to the second embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0011] Hereinafter, embodiments of a learning device and an abnormality determination system according to the present disclosure will be described with reference to the drawings. In each drawing, components denoted by the same reference numerals are identical or equivalent components, and this is common throughout the specification. Note that in each drawing, the relative dimensional relationships or shapes of each component may differ from those in reality.

[0012] Embodiment 1 1 is a control block diagram of a learning device 5 according to the first embodiment. The learning device 5 of this embodiment is a computer equipped with a processor such as a CPU or a GPU, a memory, an I / O port, and the like. The learning device 5 generates an abnormality determination model 50 for determining an abnormality of a target device. More specifically, the learning device 5 learns the relationship between the time-series operation data of the target device and the abnormal state of the target device, and generates an abnormality determination model 50 that uses the time-series operation data of the target device as input data and the abnormality determination result as output data.

[0013] As shown in FIG. 1 , the learning device 5 includes a learning data acquisition unit 51, a data preprocessing unit 52, a model generation unit 53, and a storage unit 54. The learning data acquisition unit 51, the data preprocessing unit 52, and the model generation unit 53 are functional units implemented by a processor executing a program. Alternatively, at least one of the learning data acquisition unit 51, the data preprocessing unit 52, and the model generation unit 53 may be implemented by a processing circuit such as an ASIC or FPGA. The storage unit 54 includes a volatile memory such as RAM that can temporarily store data, a nonvolatile memory such as ROM, HDD, or flash memory that can store data long-term, or both. The storage unit 54 stores the anomaly determination model 50 generated by the learning device 5.

[0014] The learning process of the abnormality determination model 50 performed by the learning device 5 will be described. The abnormality determination model 50 is a neural network that determines an abnormal state of the target equipment. In the following description, an example will be described in which the target equipment is a refrigeration cycle device that is installed in an office, a hotel, or the like and includes multiple indoor units and at least one outdoor unit. The learning device 5 of this embodiment performs two-stage learning, namely, a first learning process and a second learning process, to generate the abnormality determination model 50. FIG. 2 is a flowchart showing the flow of the learning process by the learning device 5 according to the first embodiment. FIG. 3 is a diagram explaining the flow of data in the first learning process of the first embodiment. The first learning process will be described using FIGS. 2 and 3.

[0015] 2, first, the learning data acquisition unit 51 acquires first learning data DL1 for performing the first learning process (S1). The first learning data DL1 is simulation data for development and includes a plurality of pieces of operating data representing the operating state of the refrigerant circuit of the refrigeration cycle device. The operating data is time-series measurement data measured in a certain interval (time m to time m+k) by a temperature sensor or pressure sensor provided in the refrigeration cycle device, or time-series control signal data output to each device of the refrigeration cycle device in a certain interval.

[0016] The measurement data is time-series data such as the temperature or pressure of the refrigerant discharged from the compressor, the temperature or pressure of the refrigerant drawn into the compressor, or the temperature of the refrigerant flowing through the heat exchanger. The control signal data is time-series data of control signals such as the rotation speed of the compressor of the refrigeration cycle device or the opening degree of the expansion valve. In FIG. 3, the operating data includes time-series pressure data Pi1, pressure data Pi2, and temperature data Tik of the indoor unit, and time-series pressure data Po1 and Pok of the outdoor unit. The operating data may also include data of external devices other than the target device (e.g., weather information).

[0017] Next, the data preprocessing unit 52 performs preprocessing using change points of the driving data on the first learning data DL1 acquired by the learning data acquisition unit 51 (S2). In the preprocessing, the data preprocessing unit 52 performs a grouping process (S21), a score value data conversion process (S22), and an order data generation process (S23) on each of the multiple driving data included in the first learning data DL1.

[0018] (Grouping process) Grouping processing is a process for creating integrated features by statistically processing two or more pieces of operating data from a refrigeration cycle device that show the same trend. For example, in the case of a refrigeration cycle device equipped with multiple indoor units, the operating data of multiple indoor units that are physically close to each other or multiple indoor units with the same capacity (i.e., installed in rooms of the same size) is grouped. Alternatively, multiple pieces of operating data that are correlated for a single indoor unit are grouped. In this case, "correlated" refers to similar changes over time, such as the discharge temperature of an indoor unit and the refrigerant temperature of a heat exchanger equipped in the indoor unit.

[0019] Methods for integrating multiple pieces of operating data include the arithmetic mean, geometric mean, or harmonic mean of data for the same time period.The following describes the case where the refrigeration cycle device is a housing air conditioner (HAC) consisting of two indoor units and one outdoor unit, and the time-series operating data D1(t) and D2(t) of the air outlet temperatures of the two indoor units are grouped.

[0020] In this case, the data preprocessing unit 52 obtains a waveform F1(t)={D1(t)+D2(t)} / 2 by taking an arithmetic mean of the time-series data of the discharge temperature for each time as a new feature quantity F1(t) after the data integration. Alternatively, the data preprocessing unit 52 may obtain a waveform F2(t)={D1(t)*D2(t)}^(1 / 2) by taking a geometric mean of the time as a new feature quantity F2(t). Alternatively, the data preprocessing unit 52 may obtain a waveform F3(t)=2 / {1 / D1(t)+1 / D2(t)} by taking a harmonic mean of the time as a new feature quantity F3(t).

[0021] Alternatively, the data preprocessing unit 52 may combine the feature quantities F1 to F3 calculated as described above in addition to the operating data D1 to D2 to obtain the feature quantities after data integration. The data integration method for grouping is not limited to the arithmetic mean, geometric mean, or harmonic mean, and other methods may be used. However, the following requirements must be met: First, the dimensions (units) of the data must be consistent. Specifically, temperatures must be grouped together, and pressures must be grouped together. Second, although the above example describes the integration of two series of data, D1(t) and D2(t), the integration method must also be valid for a generalized number of data, N. Third, the method must be a method of normalization by dividing by the number of data (degrees of freedom of the data), such as the geometric mean, the geometric mean, or the harmonic mean. For example, simple sum, difference, product, or division are inappropriate because the values themselves increase or decrease as the number of data increases.

[0022] By performing grouping processing as preprocessing, it is possible to process the first learning data DL1 using the same algorithm even when the configuration of the device from which the data is acquired, such as the number of connected indoor units, is different. Furthermore, by performing grouping processing on similar data, it is possible to compress the operating data without affecting the accuracy of learning. Note that grouping processing in preprocessing is not essential, and may be omitted if there are not two or more pieces of operating data that show the same trend for the target device.

[0023] (Score value data processing) The score value data processing is a process of converting the grouped operating data (feature amounts) into score value data indicating change points. Specifically, the data pre-processing unit 52 first analyzes the time-series operating data and detects change points in the operating state. A change point is a change point from the operating state when the refrigeration cycle device is operating normally, and represents a sign of an abnormal state such as a failure of any of the devices that make up the refrigeration cycle device. The data pre-processing unit 52 analyzes the operating data and detects change points, for example, using a singular spectrum transform method.

[0024] Specifically, the data preprocessing unit 52 obtains, from the time-series driving data, a history matrix X1, which is a multidimensional subspace of the reference interval, and a test matrix X2, which is a multidimensional subspace of the test space. The data preprocessing unit 52 then performs singular value decomposition on the history matrix X1 and the test matrix X2 to obtain left singular vectors of the history matrix X1 and the test matrix X2. The data preprocessing unit 52 then compares the left singular vectors of the reference interval and the test interval to determine an anomaly level α, which is the distance between the time-series data of the reference interval and the time-series data of the test interval. The anomaly level α is a value that increases as the change in the driving data increases and decreases as the change decreases. The data preprocessing unit 52 then creates a data string β that sets the anomaly level α to 1 when the anomaly level α is equal to or greater than a threshold th and to 0 when the anomaly level α is less than the threshold th, and detects points in the data string β where the value changes from 0 to 1 or from 1 to 0 as change points.

[0025] Fig. 4 is a diagram showing an example of operation data of a refrigeration cycle apparatus. Fig. 5 is a diagram showing the degree of abnormality α of the operation data shown in Fig. 4. As shown in Fig. 5, in the case of the operation data shown in Fig. 4, the degree of abnormality α becomes equal to or greater than the threshold value th at around 10 seconds, and becomes less than the threshold value at around 13 seconds. The data pre-processing unit 52 creates score value data in which the change points are set to 1 and other values are set to 0. In the examples of Figs. 4 and 5, the score value of the operation data from 0 to 9 seconds is "0", the score values of 10 and 13 seconds are "1", and the score values of 11 to 12 seconds and 14 to 30 seconds are "0".

[0026] The method of converting the score value data is not limited to the above. For example, the abnormality level α may be used as the score value data as is, or a score value may be assigned based on the trend of change (whether it is increasing or decreasing) at the change point, the amount of change, or a time constant. Furthermore, the score value is not limited to the binary values of "0" or "1," but may be a multidimensional value that increases or decreases based on the trend and amount of change, etc.

[0027] (Order data generation process) The order data generation process is a process of generating order data from score value data using the time data of each operation data as an index. The order data is data that indicates the order when multiple operation data are sorted in ascending or descending order based on the time at which a change occurred. First, the data pre-processing unit 52 acquires time data of a change point in each operation data based on the score value data. The time data is the time of the first change point in each operation data, for example, the time at which the score value data first became 1. Then, the operation data is rearranged and ordered based on the time data. For example, if the discharge pressure and the discharge temperature are acquired as operation data and the time data for the discharge pressure is earlier than the time data for the discharge temperature, order data is generated in which the discharge pressure is 1 and the discharge temperature is 2.

[0028] As described above, the preprocessing by the data preprocessing unit 52 converts the time-series driving data into preprocessed data including score value data and order data. FIG. 6 shows an example of the preprocessed data processed by the data preprocessing unit 52 of the learning device 5 according to the first embodiment. As shown in FIG. 6, the preprocessed data includes order data and score value data arranged in ascending order by the order data. The preprocessed data generated by the data preprocessing unit 52 is input to the model generation unit 53 as first input data DN1.

[0029] The operating data can be compressed by preprocessing the time-series operating data using change points and uniformly setting the data outside of the change points, i.e., the data in periods with no change, to 0. This reduces the amount of calculation required and the capacity of the memory unit 54 compared to learning from long periods of operating data as is. Furthermore, data from periods with no change is dominated by system delay times that depend on factors such as the piping length of the equipment or the outside temperature. Therefore, by excluding this data, it is possible to eliminate non-essential differences between properties and make general-purpose judgments.

[0030] Next, returning to FIG. 2, the model generation unit 53 performs a first learning process for the anomaly determination model 50 (S3). As shown in FIG. 3, the anomaly determination model 50 is a neural network composed of a large number of perceptrons. Furthermore, the anomaly determination model 50 generated in the first learning process is an autoencoder in which the number of perceptrons in the input layer is the same as the number of perceptrons in the output layer. In the first learning process, unsupervised learning is performed using first input data DN1. As shown in FIG. 3, the preprocessed first input data DN1 is input to the input layer of the neural network constituting the anomaly determination model 50, and the same data as the first input data DN1 is output to the output layer.

[0031] Between layers of the neural network, values are passed through a process based on calculations using weight coefficients and response functions. In the first learning process, the model generation unit 53 adjusts the weight coefficients so that the first input data DN1 matches the output data. The weight coefficients are estimated using the well-known backpropagation method. The error function uses, for example, the sum of squared residuals, and the response function uses, for example, a sigmoid function or a ReLU function. The abnormality determination model 50 generated by the first learning process can extract features of the first input data DN1 (driving data). As shown in FIG. 3, the layer with the smallest number of perceptrons among the intermediate layers is the layer Yc that represents the features. The input layer to the layer Yc that represents the features is the encoder layer, and the layer Yc that represents the features to the output layer is the decoder layer.

[0032] When the first learning process is completed, a second learning process is performed. FIG. 7 is a diagram illustrating the flow of data in the second learning process in the first embodiment. The second learning process will be described with reference to FIGS. 2 and 7. First, to perform the second learning process, the learning data acquisition unit 51 acquires second learning data DL2 for performing the second learning process of the abnormality determination model 50 (S4). The second learning data DL2 is simulation data for development, and includes multiple pieces of operating data representing the operating state of the refrigerant circuit of the refrigeration cycle device and the abnormal state of the refrigeration cycle device when the multiple pieces of operating data are acquired. Like the first learning data DL1, the operating data is time-series measurement data measured by a sensor provided in the refrigeration cycle device during a certain interval (time m to time m+k), or time-series control signal data output to each device of the refrigeration cycle device during a certain interval.

[0033] The abnormal state is a correct label when training the abnormality determination model 50. The abnormal state is either a flag N indicating normality, or, in the case of an abnormality, a flag E1 to Ek indicating a classification of abnormal states defined in advance by the designer of the refrigeration cycle device, or a flag Eu indicating an undefined abnormality that cannot be classified into any category. Classifications of abnormal states include, for example, high-pressure abnormality, compressor abnormality, sensor abnormality, or inverter abnormality. The learning data acquisition unit 51 acquires a plurality of second learning data DL2 in various abnormal states of various refrigeration cycle devices, which are necessary to generate the abnormality determination model 50.

[0034] Next, the data preprocessing unit 52 performs preprocessing using the change points of the driving data on the second learning data DL2 acquired by the learning data acquisition unit 51 (S5). Here, as the preprocessing, the driving data included in the second learning data DL2 is subjected to the same grouping process, score value data conversion process, and order data generation process as in the first learning process, and second input data DN2 including preprocessed data and abnormal states is obtained.

[0035] Next, the model generation unit 53 performs a second learning process on the anomaly determination model 50 (S6). In the second learning process, the encoder layer of the anomaly determination model 50 after the first learning process is maintained as is, and supervised learning using the second input data DN2 is performed on the decoder layer. As shown in Fig. 7, preprocessed data of the second input data DN2 is input to the input layer of the neural network that constitutes the anomaly determination model 50. Then, flag data indicating a normal state or, if an abnormality occurs, indicating an abnormal state is output as a determination result to the output layer.

[0036] Specifically, the weighting coefficients of the decoder layer of the anomaly determination model 50 are adjusted so that the output data when the preprocessed data is input to the anomaly determination model 50 corresponds to the anomaly state (correct label) included in the second input data DN2. The weighting coefficients are estimated using a known backpropagation method. As an example, the sum of squared residuals is used as the error function, and as an example, a sigmoid function or a ReLU function is used as the response function. The model generation unit 53 trains the anomaly determination model 50 using multiple pieces of operating data in various abnormal states and abnormal states of refrigeration cycle devices installed in various properties, acquired as the second learning data DL2.

[0037] The abnormality determination model 50 learned by the model generation unit 53 is stored in the storage unit 54 (S7). The learned abnormality determination model 50 is stored in the storage unit 34 of the abnormality determination device 3 when the abnormality determination device 3 described below is shipped, and is used to determine abnormalities in target equipment such as a refrigeration cycle device after installation in a property.

[0038] Next, an abnormality determination process using the abnormality determination model 50 generated by the learning device 5 will be described. Fig. 8 is a schematic configuration diagram of an abnormality determination system 100 according to the first embodiment. The abnormality determination system 100 in Fig. 8 comprises a refrigeration cycle apparatus 10 including one outdoor unit 1 and multiple indoor units 2, which is installed in an office or a hotel, for example, and an abnormality determination device 3. The abnormality determination device 3 determines an abnormality in the refrigeration cycle apparatus 10. The outdoor unit 1 and the multiple indoor units 2 are connected by piping and wiring such as power lines or signal lines. Furthermore, the abnormality determination device 3 is connected to the outdoor unit 1 and the multiple indoor units 2 so as to be able to communicate with them via wire or wireless communication.

[0039] Fig. 9 is a refrigerant circuit diagram of a refrigeration cycle apparatus 10 according to the first embodiment. As shown in Fig. 9, the outdoor unit 1 includes a compressor 11, a flow path switching valve 12, an outdoor heat exchanger 13, an expansion valve 14, an outdoor fan 15, and a control device 16. Each indoor unit 2 includes an indoor heat exchanger 21 and an indoor fan 22. The compressor 11, the flow path switching valve 12, the outdoor heat exchanger 13, the expansion valve 14, and the indoor heat exchanger 21 are connected by piping to form a refrigerant circuit. The refrigerant circulating in the refrigeration cycle apparatus 10 is, for example, a natural refrigerant such as carbon dioxide, hydrocarbon, or helium; a chlorine-free refrigerant such as HFC410A or HFC407C; or a fluorocarbon refrigerant such as R22 or R134a.

[0040] The compressor 11 draws in and compresses a low-pressure gas refrigerant, and discharges it as a high-pressure gas refrigerant. As the compressor 11, for example, a reciprocating, rotary, scroll, or screw compressor is used. The operating frequency of the compressor 11 is controlled by a control device 16.

[0041] The flow path switching valve 12 switches between a cooling operation in which the outdoor heat exchanger 13 functions as a condenser and a heating operation in which the outdoor heat exchanger 13 functions as an evaporator. The flow path switching valve 12 is, for example, a four-way valve, and its switching is controlled by the control device 16. During the cooling operation, the flow path switching valve 12 switches so that the refrigerant discharged from the compressor 11 flows into the outdoor heat exchanger 13, as shown by the solid line in Fig. 1. During the heating operation, the flow path switching valve 12 switches so that the refrigerant discharged from the compressor 11 flows into the indoor heat exchanger 21, as shown by the dashed line in Fig. 1.

[0042] The outdoor heat exchanger 13 is, for example, a fin-tube heat exchanger, and exchanges heat between the refrigerant flowing inside the circular or flat tubes and the air supplied by the outdoor fan 15. The outdoor heat exchanger 13 is disposed between the flow path switching valve 12 and the expansion valve 14. The outdoor heat exchanger 13 functions as an evaporator during heating operation and as a condenser during cooling operation.

[0043] The expansion valve 14 is disposed between the outdoor heat exchanger 13 and the indoor heat exchanger 21, and reduces the pressure of the refrigerant to expand it. The expansion valve 14 is, for example, an electronic expansion valve whose opening degree is variable. The opening degree of the expansion valve 14 is controlled by the control device 16.

[0044] The outdoor fan 15 draws in outdoor air, passes it through the outdoor heat exchanger 13, and blows it outside. The outdoor fan 15 is, for example, a propeller fan, turbo fan, sirocco fan, or crossflow fan driven by a motor. The airflow rate of the outdoor fan 15 is controlled by a control device 16.

[0045] The indoor heat exchanger 21 is, for example, a fin-tube type heat exchanger, and exchanges heat between the refrigerant flowing inside the circular or flat tubes and the air blown by the indoor fan 22. The indoor heat exchanger 21 is disposed between the expansion valve 14 and the flow path switching valve 12. The indoor heat exchanger 21 functions as a condenser during heating operation, and as an evaporator during cooling operation.

[0046] The indoor fan 22 draws in indoor air, passes it through the indoor heat exchanger 21, and blows it out into the room. The indoor fan 22 is, for example, a propeller fan, turbo fan, sirocco fan, or crossflow fan driven by a motor. The airflow rate of the indoor fan 22 is controlled by the control device 16.

[0047] The refrigeration cycle apparatus 10 also includes a plurality of sensors for measuring operating data. Specifically, the refrigeration cycle apparatus 10 includes a discharge temperature sensor 41 that measures the temperature of the refrigerant discharged from the compressor 11, and an intake temperature sensor 42 that measures the temperature of the refrigerant sucked into the compressor 11. The discharge temperature sensor 41 is provided in the discharge piping of the compressor 11, and the intake temperature sensor 42 is provided in the suction piping of the compressor 11.

[0048] The refrigeration cycle device 10 also includes a first refrigerant temperature sensor 43 and a second refrigerant temperature sensor 44 that measure the temperature of the refrigerant flowing into or out of the outdoor heat exchanger 13. The first refrigerant temperature sensor 43 measures the temperature of the refrigerant flowing into the outdoor heat exchanger 13 during cooling operation, and the second refrigerant temperature sensor 44 measures the temperature of the refrigerant flowing out of the outdoor heat exchanger 13 during cooling operation.

[0049] The refrigeration cycle device 10 also includes a third refrigerant temperature sensor 45 and a fourth refrigerant temperature sensor 46 that measure the temperature of the refrigerant flowing into or out of the indoor heat exchanger 21. The third refrigerant temperature sensor 45 measures the temperature of the refrigerant flowing into the indoor heat exchanger 21 during cooling operation, and the fourth refrigerant temperature sensor 46 measures the temperature of the refrigerant flowing out of the indoor heat exchanger 21 during cooling operation.

[0050] Each of the above temperature sensors is configured by a thermistor, for example. The refrigeration cycle apparatus 10 may also include a temperature sensor or pressure sensor other than the above temperature sensors. For example, the refrigeration cycle apparatus 10 may also include an indoor temperature sensor provided in the indoor unit 2 to measure the temperature of the indoor air, or an outdoor air temperature sensor located outside the housing of the outdoor unit 1 to measure the temperature of the air outside the room where the outdoor unit 1 is installed. Furthermore, at least one of the discharge temperature sensor 41, the suction temperature sensor 42, and the first refrigerant temperature sensor 43 to the fourth refrigerant temperature sensor 46 may also be a pressure sensor that measures the pressure of the refrigerant.

[0051] The control device 16 is a computer including, for example, a processor such as a CPU, a memory, an I / O port, etc. The control device 16 controls the operation of the refrigeration cycle device 10 based on instructions from a user input via a remote controller (not shown) of the refrigeration cycle device 10 and measurement results from each temperature sensor. Specifically, the control device 16 controls the operating frequency of the compressor 11, the switching of the flow path switching valve 12, the opening degree of the expansion valve 14, and the rotation speeds of the outdoor fan 15 and the indoor fan 22 based on the setting of the operation mode and the set temperature input by the user and the measurement results from each temperature sensor.

[0052] Returning to FIG. 8 , the abnormality determination device 3 is a computer equipped with a processor such as a CPU or GPU, a memory, an I / O port, and the like. As shown in FIG. 8 , the abnormality determination device 3 includes an operating data acquisition unit 31, a data preprocessing unit 32, an abnormality determination unit 33, and a storage unit 34. The operating data acquisition unit 31, the data preprocessing unit 32, and the abnormality determination unit 33 are functional units implemented by a processor executing a program. Alternatively, at least one of the operating data acquisition unit 31, the data preprocessing unit 32, and the abnormality determination unit 33 may be implemented by a processing circuit such as an ASIC or FPGA. The storage unit 34 includes a volatile memory such as RAM that can temporarily store data, a nonvolatile memory such as ROM, HDD, or flash memory that can store data for a long period of time, or both. The storage unit 34 stores the trained abnormality determination model 50 generated by the learning device 5.

[0053] An example of abnormality determination for the refrigeration cycle apparatus 10 will be described. Consider a case where the indoor unit 2 of the refrigeration cycle apparatus 10 is performing cooling operation and an abnormality called a high-pressure abnormality occurs in which the discharge pressure of the compressor 11 exceeds a preset threshold. The high-pressure abnormality occurs, for example, in the following case. (1) In the environment where the outdoor unit 1 was installed, wind flowed in in the direction opposite to the direction of airflow from the outdoor fan 15 of the outdoor unit 1. (2) The heat exchange capacity of the outdoor heat exchanger 13 of the outdoor unit 1 has decreased. (3) In order to increase the heat exchange amount of the outdoor heat exchanger 13 of the outdoor unit 1, the rotation speed of the outdoor fan 15 is controlled to increase. (4) Controlling the rotation speed of the compressor 11 to increase the amount of refrigerant circulating. (5) The discharge pressure of the compressor 11 exceeds a preset threshold, causing the compressor 11 to stop or fail to provide sufficient cooling capacity.

[0054] In addition to the above, (1) may be caused by environmental factors such as a rise in outdoor air temperature, direct sunlight on the outdoor unit 1, or the outdoor unit 1 being installed on concrete or other surfaces that easily store heat. It may also be caused by an installation environment factor known as a short circuit, in which warm air discharged from the outdoor unit 1 air outlet is sucked back into the air inlet when the outdoor unit 1 is close to a wall in the installation location. The above abnormalities can be determined based on control signal data such as the rotation speeds of the outdoor fan 15 and the indoor fan 22, the rotation speed of the compressor 11, and the number of pulses that adjust the opening of the expansion valve 14, as well as measurement data such as the discharge temperature, intake temperature, and first to fourth refrigerant temperatures of the compressor 11.

[0055] Fig. 10 is a flowchart showing the flow of abnormality determination processing by the abnormality determination device 3 according to embodiment 1. Fig. 11 is a diagram illustrating the flow of data in the abnormality determination processing of embodiment 1. The abnormality determination processing of the refrigeration cycle device 10 by the abnormality determination device 3 will be described with reference to Figs. 10 and 11. First, the operating data acquisition unit 31 acquires a plurality of operating data of the refrigeration cycle device 10 (S11).

[0056] The operating data represents the operating state of the refrigerant circuit of the refrigeration cycle apparatus 10. The operating data is time-series measurement data measured over a certain period by a temperature sensor or pressure sensor provided in the refrigeration cycle apparatus 10, or time-series control signal data output to each device of the refrigeration cycle apparatus 10 over a certain period. The measurement data is, for example, the discharge temperature Te measured by the discharge temperature sensor 41, the suction temperature Ts measured by the suction temperature sensor 42, and the first refrigerant temperature Tr1 to the fourth refrigerant temperature Tr4 measured by the first refrigerant temperature sensor 43 to the fourth refrigerant temperature sensor 46. The control signal data is time-series data such as the number of pulses for adjusting the rotation speed of the compressor 11 or the opening of the expansion valve 14. The operating data may also include data from external devices other than the refrigeration cycle apparatus 10 (for example, weather information).

[0057] Next, the data pre-processing unit 32 performs pre-processing on the driving data acquired by the driving data acquisition unit 31 (S12). Specifically, like the data pre-processing unit 52 of the learning device 5, the data pre-processing unit 32 performs grouping processing (S21), score value data conversion processing (S22), and order data generation processing (S23) as pre-processing. The grouping processing, score value data conversion processing, and order data generation processing in the data pre-processing unit 32 are the same as the grouping processing, score value data conversion processing, and order data generation processing performed by the data pre-processing unit 52 of the learning device 5. Through the pre-processing by the data pre-processing unit 32, the time-series driving data is converted into pre-processed data consisting of score value data and order data.

[0058] Next, the abnormality determination unit 33 performs an abnormality determination using the trained abnormality determination model 50 stored in the memory unit 54 (S13). The abnormality determination unit 33 inputs the preprocessed data preprocessed by the data preprocessing unit 52 into the abnormality determination model 50 and obtains an abnormality determination result as output data. The abnormality determination result is flag data indicating a normal or classified abnormal state. The abnormality determination device 3 transmits the obtained determination result to the refrigeration cycle apparatus 10 or a management device that manages the refrigeration cycle apparatus 10. The refrigeration cycle apparatus 10 or the management device displays the determination result on a display device and notifies the user. Alternatively, the abnormality determination device 3 may be equipped with a display device and display the determination result.

[0059] As described above, in this embodiment, a large amount of operating data of the refrigeration cycle apparatus 10 used for learning or abnormality detection is preprocessed using change points to convert it into preprocessed data consisting of score value data and order data. This allows the operating data used for learning the abnormality detection model 50 and abnormality detection using the abnormality detection model 50 to be compressed, thereby reducing the load on the apparatus that performs the learning process for the abnormality detection model 50 and the abnormality detection process using the abnormality detection model 50. As a result, high-performance and large-capacity apparatus is not required, preventing increases in apparatus and system costs. Furthermore, improved processing speed allows for rapid learning of the abnormality detection model 50 and abnormality detection of target equipment. This reduces downtime due to malfunctions and enables the refrigeration cycle apparatus 10 to maintain comfort when used in a living space, preserve food when used in a refrigerator / freezer, and ensure stable operation of the computer equipment when used for cooling purposes such as in a server room.

[0060] Embodiment 2 A second embodiment will be described. The second embodiment differs from the first embodiment in that the abnormality determination model 50 is re-learned in accordance with the result of the abnormality determination process. The configuration and learning process of the learning device 5 and other configurations of the abnormality determination system 100 are the same as those of the first embodiment.

[0061] 12 is a schematic configuration diagram of an abnormality determination system 100A according to embodiment 2. The abnormality determination system 100A of this embodiment comprises a refrigeration cycle apparatus 10 consisting of one outdoor unit 1 and multiple indoor units 2, and an abnormality determination device 3A. The configurations of the outdoor unit 1 and the indoor units 2 are the same as those in embodiment 1.

[0062] 12, the abnormality determination device 3A of this embodiment includes an operating data acquisition unit 31, a data preprocessing unit 32, an abnormality determination unit 33, and a storage unit 34, which are the same as those of the first embodiment, as well as a relearning unit 35. The relearning unit 35 is a functional unit that is realized by a processor executing a program. Alternatively, the relearning unit 35 may be realized by a processing circuit such as an ASIC or an FPGA.

[0063] As described in the first embodiment, the abnormality determination model 50 used in the abnormality determination device 3A is trained by the learning device 5 using learning data (simulation data for development). The capacity and number of connected indoor units 2 differ for each property where the refrigeration cycle device 10 is installed. Therefore, the dimensions and ranges of the time-series operating data to be analyzed and determined differ between different properties, making it difficult to share learning data between the properties. Furthermore, because the operating data obtained after installation is biased, it is not possible to obtain sufficient data for training the abnormality determination model 50 from scratch using data from a single property. Therefore, in the abnormality determination device 3A of this embodiment, the re-learning unit 35 re-learns the abnormality determination model 50 stored in the memory unit 34 based on the determination results of the abnormality determination unit 33 and feedback from users.

[0064] 13 is a flowchart showing the flow of an abnormality determination process by the abnormality determination device 3A according to the second embodiment. The processes of steps S11 to S13 in the abnormality determination process of this embodiment are the same as the processes of steps S11 to S13 in the abnormality determination process of the first embodiment. When an abnormality determination result is obtained in step S13, the re-learning unit 35 determines whether the abnormality determination result is an erroneous determination (S14). Here, for example, if an abnormality determination result is obtained but there is no abnormality, the user may input feedback indicating that the abnormality was normal via an operation unit (not shown) of the abnormality determination device 3A. When the feedback from the user differs from the abnormality determination result, the re-learning unit 35 determines that the abnormality determination is an erroneous determination in step S14.

[0065] If the determination result is not an erroneous determination (S14: NO), there is no problem with the abnormality determination model 50 and re-learning is not necessary, so the abnormality determination process is terminated. On the other hand, if the determination result is an erroneous determination (S14: YES), the re-learning unit 35 performs a re-learning process for the abnormality determination model 50 using feedback from the user and the driving data at the time of the abnormality determination (S15). Specifically, pre-processed data obtained by pre-processing the driving data at the time of the abnormality determination is input to the abnormality determination model 50, and a weighting coefficient is estimated so that the output data is the feedback from the user.

[0066] Here, the re-learning unit 35 does not change the weighting coefficients between all layers of the abnormality determination model 50, but only changes the weighting coefficients of the decoder layer, i.e., the intermediate layers from the layer Yc (FIG. 11) representing the feature quantities. The layer Yc representing the feature quantities is a layer that represents the feature quantities of the driving data, and is the layer that has the fewest number of perceptrons and is closest to the output layer in the abnormality determination model 50. In the trained abnormality determination model 50, the layers from the input layer to the layer Yc representing the feature quantities are kept confidential, and are configured so that they cannot be rewritten by re-learning.

[0067] The weight coefficients used in re-learning are estimated using the well-known backpropagation method. The error function is, for example, the sum of squared residuals, and the response function is, for example, a sigmoid function or a ReLU function. By limiting the weight coefficients changed during re-learning during operation to the layers after the layer Yc, which represents the feature values, the number of weight coefficients to be re-learned can be reduced compared to re-learning all of the weight coefficients of the neural network. This reduces the load of re-learning and enables appropriate re-learning even with a small dataset obtained during operation.

[0068] The re-learning unit 35 stores the re-learned abnormality determination model 50 in the storage unit 34. As a result, the abnormality determination model 50 is updated (S16).

[0069] As described above, in the abnormality determination device 3A of this embodiment, the abnormality determination model 50 is updated in accordance with the environment after the refrigeration cycle apparatus 10 is installed or the way the user uses the refrigeration cycle apparatus 10 at the installation site. This makes it possible to obtain abnormality determination results with higher accuracy than at the time of shipment. Note that the updating of the abnormality determination model 50 is not limited to when an erroneous determination is made as described above, but may also be performed periodically. For example, at a certain time (m+k), the actual abnormal state may be confirmed for each of a plurality of determination results from past times (m) to (m+k-1), and the abnormality determination model 50 may be re-learned based on the confirmation results.

[0070] The above is a description of the embodiments, but the present disclosure is not limited to the above embodiments and can be modified in various ways without departing from the spirit and scope of the present disclosure. Furthermore, the present disclosure includes any combination of the configurations described in the above embodiments. For example, the affiliation of each part of the learning device 5 and the abnormality determination device 3 is not limited to the example of the above embodiments. For example, the control device 16 of the outdoor unit 1 may have some or all of the functional parts of the abnormality determination device 3 and some or all of the functional parts of the learning device 5. Alternatively, the abnormality determination device 3 may have some or all of the functional parts of the learning device 5. When a single device has the functions of the abnormality determination device 3 and the learning device 5, learning, relearning, and abnormality determination of the abnormality determination model 50 can be switched by operating the device. Alternatively, each functional part of the abnormality determination device 3 or each functional part of the learning device 5 may be implemented by multiple computers. Furthermore, the abnormality determination model 50 may be stored in a cloud storage device.

[0071] Furthermore, in the above embodiment, the refrigeration cycle apparatus 10 is described as having a plurality of indoor units 2, but the present disclosure is not limited thereto. The refrigeration cycle apparatus 10 may be a room air conditioner having one outdoor unit 1 and one indoor unit 2. Furthermore, in the above embodiment, the refrigeration cycle apparatus 10 is described as an air conditioner, but the refrigeration cycle apparatus 10 may be a heat pump chiller, a cooling-only machine without a cooling / heating switchable unit, a refrigeration unit for cooling a refrigerated warehouse, or a water heater. Furthermore, the target device for which an abnormality is determined by the abnormality determination device 3 is not limited to the refrigeration cycle apparatus 10, but may be any device for which an abnormality can be determined based on time-series operating data.

[0072] Furthermore, the learning method of the abnormality determination model 50 in the learning device 5 is not limited to the example of the above embodiment. For example, the learning device 5 may perform only the second learning process to generate the abnormality determination model 50 in which a plurality of time-series driving data are used as input data and an abnormality determination result is used as output data. [Explanation of symbols]

[0073] 1 outdoor unit, 2 indoor unit, 3, 3A abnormality determination device, 5 learning device, 10 refrigeration cycle device, 11 compressor, 12 flow path switching valve, 13 outdoor heat exchanger, 14 expansion valve, 15 outdoor fan, 16 control device, 21 indoor heat exchanger, 22 indoor fan, 31 operating data acquisition unit, 32 data pre-processing unit, 33 abnormality determination unit, 34 memory unit, 35 relearning unit, 41 discharge temperature sensor, 42 suction temperature sensor, 43 first refrigerant temperature sensor, 44 second refrigerant temperature sensor, 45 third refrigerant temperature sensor, 46 fourth refrigerant temperature sensor, 50 abnormality determination model, 51 learning data acquisition unit, 52 data pre-processing unit, 53 model generation unit, 54 memory unit, 100, 100A abnormality determination system.

Claims

1. a learning data acquisition unit that acquires learning data including a plurality of time-series operation data of the target device; a data preprocessing unit that performs preprocessing of the driving data using the change points of the driving data; a model generation unit that uses the preprocessed operating data to generate an abnormality determination model in which the plurality of time-series operating data of the target device are input data and an abnormality determination result of the target device is output data.

2. The learning device according to claim 1 , wherein the data preprocessing unit detects the change points in the driving data and converts the driving data into different score value data depending on the change points.

3. The data preprocessing unit generating order data of the plurality of driving data based on the time of the first change point in the plurality of driving data; The learning device according to claim 2 , wherein the score value data and the order data are used as the input data for the anomaly determination model.

4. The data preprocessing unit performing a grouping process of integrating two or more pieces of driving data showing the same tendency among the plurality of driving data; The learning device according to claim 2 or 3, wherein the change point is detected from the driving data after the grouping process.

5. an operation data acquisition unit that acquires a plurality of time-series operation data of the target device; a data preprocessing unit that performs preprocessing of the driving data using the change points of the driving data; an abnormality determination unit that determines an abnormality in the target device using a trained abnormality determination model that uses the preprocessed operating data as input data and an abnormality determination result of the target device as output data.

6. The abnormality determination system according to claim 5 , further comprising a re-learning unit that re-learns the abnormality determination model when the determination result of the abnormality determination unit is an erroneous determination or periodically.

7. The abnormality determination system according to claim 6 , wherein the re-learning unit re-learns intermediate layers subsequent to a layer representing a feature amount in the abnormality determination model.

Images