Information processing device, model generation method

Abstract

To improve the accuracy of predictions made by normal prediction models. [Solution] This disclosure relates to an information processing device that generates a normal prediction model for predicting the operation data output by a device when it is functioning normally. The device obtains a provisional normal prediction model 53 generated using training data containing multiple records, where operation data collected from the device at the same time is treated as one record. The operation data includes a first part corresponding to an objective variable, which is information about temperature or an actuator, and a second part corresponding to explanatory variables other than the objective variable. The control unit calculates the degree of deviation between the predicted value 54 output by the provisional normal prediction model for the input of the second part of any record and the first part of the same record as the second part 55, for each record in the training data. Based on statistical information of multiple deviations, the control unit identifies valid records and generates a normal prediction model using the valid records.

Description

Technical Field

[0001] The present invention relates to an information processing apparatus and a model generation method.

Background Art

[0002] A technique is known in which a predicted value is calculated using a normal prediction model, and a failure of a device is detected based on the degree of deviation between the predicted value and the measured value. The normal prediction model is a model learned using learning data when the device is normal, and is a model for predicting data output by the device when it is normal. Therefore, when the degree of deviation is large, it can be detected that a sensor or the like that outputs the measured value has failed.

[0003] Patent Document 1 discloses a technique for determining a specific type of failure such as electrical intermittency or freezing of a sensor by recognizing a pattern in which sensor data deviates from a normal distribution.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] However, in the conventional technology, there is a possibility that the accuracy of the predicted value output by the normal prediction model may decrease. Although outliers are removed from the learning data for generating the normal prediction model, in the conventional technology, data output by a temperature sensor or an actuator having an abnormality is not removed, so there is a possibility that the accuracy of the predicted value output by the normal prediction model may decrease.

[0006] The present disclosure provides a technique for improving the accuracy of a predicted value by a normal prediction model.

Means for Solving the Problems

[0007] The first aspect of this disclosure is, An information processing device that generates a normal prediction model for predicting the operating data that a device outputs when it is functioning normally, It has a control unit, A hypothetical normal prediction model is generated using training data containing multiple records, with operating data collected simultaneously from the aforementioned equipment treated as one record, and the following: The aforementioned operating data is acquired, The aforementioned operating data includes a first part corresponding to a target variable, which is information about temperature or the actuator, and a second part corresponding to an explanatory variable, which is information about temperature or the actuator other than the target variable. The control unit calculates the degree of deviation between the predicted value output by the pseudo-normal prediction model for the input of the second portion of any record and the first portion of the same record, for each record of the training data. Based on statistical information on multiple deviations, valid records are identified. The normal prediction model is generated using the valid records mentioned above.

[0008] According to the first aspect of this disclosure, the accuracy of the predicted values ​​by the normal prediction model can be improved.

[0009] A second aspect of this disclosure is an information processing apparatus described in the first aspect, The control unit, Remove the records containing the first and second portions used to calculate the degree of deviation identified as the outliers, The normal prediction model is generated using the training data from which the aforementioned records have been removed.

[0010] A third aspect of this disclosure is an information processing apparatus described in the second aspect, The operating data collected from the aforementioned equipment at the same time is further comprised of, for each record, a third portion corresponding to a second objective variable which is information about temperature or actuators, and a fourth portion corresponding to a second explanatory variable which is information about temperature or actuators other than the second objective variable. The control unit generates the normal prediction model using the training data from which the records have been removed, A second normal prediction model is generated that associates the third part with the fourth part.

[0011] A fourth aspect of this disclosure is an information processing device as described in the second aspect, The aforementioned statistical information consists of the mean or median of multiple deviations and the deviation from the mean or median. The control unit removes records that include the first and second portions used to calculate the deviation, where the deviation from the mean or median is greater than or equal to a threshold.

[0012] A fifth aspect of this disclosure is an information processing device as described in the second aspect, The statistical information consists of quartiles of multiple degrees of deviation. The control unit uses N times the IQR (Interquotile Range) as a threshold, The control unit removes records that include the first and second portions used to calculate the deviation, which are further than a threshold from the first or third quartile.

[0013] A sixth aspect of this disclosure is an information processing apparatus described in the fourth or fifth aspect, If the number of records included in the aforementioned training data is less than a certain amount, The control unit reduces the threshold value to a smaller value than when the number of records is not less than a certain value.

[0014] A seventh aspect of this disclosure is an information processing device described in any of the first to sixth aspects, The control unit, Excluding records with transient data and error codes attached from the learning data, a provisional normal prediction model is generated.

[0015] An eighth aspect of the present disclosure is an information processing apparatus according to any one of the first to seventh aspects, where the control unit inputs the second part corresponding to the explanatory variable acquired from the device during operation of the device into the normal prediction model to calculate a predicted value, and detects a failure of a component that output the first part based on a degree of divergence between the first part corresponding to the target variable acquired from the device simultaneously with the explanatory variable and the predicted value.

[0016] A ninth aspect of the present disclosure is an information processing apparatus according to any one of the first to eighth aspects, where the device is an air conditioner.

[0017] A tenth aspect of the present disclosure is a model generation method for an information processing apparatus that generates a normal prediction model for predicting operation data output by a device when the device is normal, the information processing apparatus generating the normal prediction model, where the information processing apparatus has a control unit and a provisional normal prediction model generated using learning data including a plurality of records with operation data collected simultaneously from the device as one record, the control unit acquires the operation data, the operation data includes a first part corresponding to a target variable that is information related to temperature or an actuator and a second part corresponding to an explanatory variable that is information related to temperature or an actuator other than the target variable, the control unit calculates, for each record of the learning data, a degree of divergence between a predicted value output by the provisional normal prediction model for an input of the second part of an arbitrary record and the first part of the record having the same second part, identifies valid records based on statistical information of a plurality of degrees of divergence, The normal prediction model is generated using the valid records mentioned above.

[0018] According to the tenth aspect of this disclosure, the accuracy of the predicted values ​​by the normal prediction model can be improved. [Brief explanation of the drawing]

[0019] [Figure 1] This diagram illustrates the process for detecting a malfunction in an air conditioner, as an example of equipment. [Figure 2] This is a system configuration diagram of an example of an equipment management system. [Figure 3] This diagram illustrates an example of the arrangement of fault detection devices and learning devices in an equipment management system. [Figure 4] This is a hardware configuration diagram of an example edge device. [Figure 5] This is a hardware configuration diagram of an example server device. [Figure 6] This is a functional block diagram of an example of a learning device. [Figure 7] This is a functional block diagram of an example of a fault detection device. [Figure 8] This figure shows a list of example driving data. [Figure 9] This figure shows an example of a hypothetical normal prediction model or normal prediction model constructed using a neural network. [Figure 10] This diagram illustrates the outlier removal process during the learning phase. [Figure 11] This diagram schematically illustrates transient data. [Figure 12] This figure shows an example of the distribution of the degree of deviation. [Figure 13] This diagram illustrates a method for detecting outliers using quartiles. [Figure 14] This figure shows an example of the relationship between the number of training data points and the threshold. [Figure 15] This is an example flowchart illustrating the process by which a learning device generates a normal prediction model. [Figure 16]This is a flowchart illustrating the process by which a fault detection device detects a malfunction in an air conditioner. [Figure 17] This figure shows a graph overlaid with measured and predicted values ​​over time. [Modes for carrying out the invention]

[0020] The following describes an information processing device and a model generation method performed by the information processing device as an example of a form for implementing this disclosure.

[0021] <Overview> Figure 1 illustrates the process for detecting a malfunction in an air conditioner, as an example of equipment. (1) Preprocessing is performed on the operating data acquired from the air conditioner. Preprocessing is a process that removes, for example, transient data and operating data with attached error codes. Error codes are numbers or symbols that are displayed when the air conditioner does not operate normally or when a malfunction occurs in the air conditioner, and they indicate error identification information. (2) The normal prediction model (assumed to be pre-generated in the explanation of Figure 1) takes at least some of the explanatory variables from the operating data as input and outputs a predicted value. This predicted value is a prediction of one of the measured values ​​in the operating data. Therefore, the normal prediction model is a model that predicts any measured value when the air conditioner is in a normal state. (3) The degree of deviation between the predicted value and the measured value, which is one of the operating data, is calculated. This measured value is the operating data that was used as the target variable when the normal prediction model was generated. (4) Whether or not the air conditioner is malfunctioning is determined based on whether the degree of deviation is above a threshold. In other words, if the sensor or actuator that outputs the measured value is malfunctioning, the deviation between the predicted value and the measured value will be large. This is because the normal prediction model is a model that has been trained using training data from when the equipment is functioning normally, and is a model that predicts the data that the equipment will output when it is functioning normally. For this reason, if the degree of deviation is above a threshold, it can be determined that the sensor or actuator that outputs the measured value is malfunctioning. (5) If it is determined that the air conditioner has malfunctioned, an alert will be sent to the person in charge.

[0022] It is preferable that the training data used to generate a normal operation prediction model does not include operating data output by faulty sensors or actuators. Therefore, when preparing training data to generate a normal operation prediction model, unreliable data is removed, but currently some of this removal is done manually, which is time-consuming.

[0023] Therefore, in this embodiment, effective operating data is identified by removing outliers from the training data using the following procedure. 1. The learning device generates a hypothetical normal prediction model using preprocessed training data. The hypothetical normal prediction model is, by definition, a hypothetical normal prediction model. This hypothetical normal prediction model is generated using training data that may contain outliers. 2. The learning device calculates the degree of discrepancy between the predicted value and the actual value using the hypothetical normal prediction model. 3. The learning device calculates, as an example, the mean and standard deviation (σ) of the degree of deviation. 4. The learning device removes measured values ​​that are, for example, more than ±3σ from the mean from the training data. When removing these values, it is preferable to delete not only the driving data used to generate the hypothetical normal prediction model, but also all driving data (1 record) collected at the same time. 5. The learning device generates a normal prediction model using the training data from which outliers have been removed.

[0024] Thus, the learning device of this embodiment only requires the removal of outliers using statistical information such as the mean and σ, so even personnel with limited knowledge of air conditioners can remove outliers. Furthermore, it can detect failures in actuators, compressors, expansion valves, and other components besides thermistor sensors.

[0025] <About Terminology> A model is a mathematical representation or algorithm that learns patterns and rules from data and uses that knowledge to make predictions and decisions about new data. Specifically, a model is built using training data, capturing the features and patterns of that training data, and is designed to identify similar features in unknown data and output the correct results.

[0026] A hypothetical normal prediction model is a normal prediction model generated at the preprocessing stage. The normal prediction model is a model generated after removing operating data that is judged to be outliers using the hypothetical normal prediction model. Therefore, the normal prediction model is a model that learns the correspondence between any first and second data points from the operating data output by the equipment under normal conditions, using training data from when the equipment is functioning normally.

[0027] "Operation" refers to the movement or operation of a machine, and "operation data" refers to one or more data points obtained as a result of the machine's operation. Operation data is collected from equipment at the same time and may refer to multiple operation data points. Multiple operation data points collected from equipment at the same time are referred to as one record. In this disclosure, a normal prediction model is generated using training data having multiple records. "At the same time" does not necessarily mean the exact same time; it is sufficient if the data is collected at the same interval. For example, if data is collected at a fixed time every hour, operation data collected within one hour of that time can be considered operation data collected at the same time.

[0028] The first data set is the operational data corresponding to the dependent variable from a single record of operational data collected from the equipment at the same time, while the second data set is the operational data corresponding to the independent variable. Although there is only one dependent variable in the normal prediction model or the hypothetical normal prediction model, multiple normal prediction models or hypothetical normal prediction models are generated. Therefore, the dependent variable differs for each normal prediction model or hypothetical normal prediction model.

[0029] A record refers to one row of data when data is arranged in a two-dimensional table. One column of data is a column. In this disclosure, different types of operational data are arranged in the row direction, and operational data is arranged chronologically in the column direction.

[0030] Statistical information is a quantitative description of the characteristics of multiple data points, such as the mean, median, standard deviation, variability, and distribution. In this embodiment, any statistical information that can identify outliers is sufficient.

[0031] Removal means to take away or remove. In this disclosure, it means not to include it in the training data. Removal may also mean deletion.

[0032] <System configuration of the equipment management system> Referring to Figure 2, the system configuration of the equipment management system 100 will be described. Figure 2 is a diagram showing an example of the system configuration of the equipment management system 100.

[0033] The equipment management system 100 provides a variety of IoT-based services to users, from administrators to general users, by enabling communication between various devices 30 such as air conditioners and lighting and a server device 60 on the cloud side via a network N. The edge devices 10, devices 30, sensor switches 31, and user terminals 70 are mainly located on the customer side, while the server device 60 is located in a cloud such as a data center or the internet.

[0034] Equipment 30 refers to all devices that consume power, such as air conditioners, security equipment, heat source equipment, fire alarms, AHUs (air handling units), electricity meters, and lighting. Equipment 30 may also include other devices. Sensor switches 31 include various sensors, lamps, relays, etc. Actuators may also be included in sensor switches 31. Equipment 30 and sensor switches 31 are connected to the edge device 10 so as to be able to communicate via a dedicated cable or a network such as a LAN. Equipment 30 and sensor switches 31 may also be connected to the edge device 10 so as to be able to communicate via wireless communication.

[0035] The equipment 30 and sensor switches 31 are controlled by the edge device 10. In other words, the edge device 10 performs the necessary operations on the equipment 30 and sensor switches 31 to suit the purpose of the equipment 30 and sensor switches 31. The content of the control varies depending on the type of equipment 30 and sensor switches 31, but for example, if the equipment 30 is an air conditioner, it may include all control related to the functions of the air conditioner, such as the cooling / heating mode, set temperature, airflow, humidity, and airflow direction that can be set on an air conditioner. In addition, the control may include operating modes such as a pre-season inspection mode, microcontroller reset, operation stop, and function substitution.

[0036] Device 30 collects operational data specific to device 30 and transmits it to the edge device 10 mainly periodically. Periodically means, for example, once every minute, once every 10 minutes, once every 60 minutes, etc., but this can be set by the user or the server device 60. Also, upon request from the edge device 10 or the user terminal 70, device 30 can transmit operational data to the edge device 10. The operational data varies depending on the device 30, but for example, in the case of an air conditioner, it may include sensor values ​​detected by sensors, actuator current values ​​and output values, compressor rotation speed, expansion valve opening, refrigerant high pressure, low pressure, refrigerant temperature, fan rotation speed, and microcontroller CPU temperature.

[0037] Furthermore, if device 30 detects an abnormality, it sends an abnormality code to the edge device 10. Device 30 that detected the abnormality stops operation. The edge device 10 sends the abnormality code to the server device 60. The processing of the edge device 10 for sensor switches 31 is the same. Sensor switches 31 mainly periodically send information about themselves to the edge device 10 and send abnormality codes.

[0038] Furthermore, the device 30 may have a normal prediction model. In this case, the device 30 can use the normal prediction model to detect failures in sensors, actuators, etc., and transmit an abnormality code to the edge device 10.

[0039] The edge device 10 is a controller that controls the equipment 30 and sensor switches 31. The edge device 10 has the functions of a control device that controls the equipment 30 and sensor switches 31, an information processing device that processes operating data, etc., and a communication device that communicates with the server device 60. For example, the edge device 10 transmits various information from the equipment 30 to the server device 60 and receives instructions from the server device 60 according to the information. Alternatively, the edge device 10 can receive instructions from the server device 60 even if it does not transmit any information to the server device 60 (for example, if there are instructions from the user terminal 70 to the server device 60). The edge device 10 converts the instructions into appropriate instructions according to the models of the equipment 30 and sensor switches 31 and transmits them to the equipment 30 and sensor switches 31.

[0040] The edge device 10 may also have a normal prediction model. In this case, the edge device 10 can use the normal prediction model to detect failures in sensors, actuators, etc., and send an error code to the server device 60.

[0041] A server device 60 is one or more information processing devices. Although Figure 2 shows one server device 60, the server devices 60 may be divided into several units according to their functions. Alternatively, the functions of the server devices 60 may be consolidated into a single information processing device. Furthermore, multiple server devices 60 with the same functions may be provided, and these multiple server devices 60 may communicate with each other to process data, similar to a server cluster.

[0042] The server device 60 receives various information transmitted from the edge device 10 via the network N and generates necessary instructions. For example, in response to an abnormal code from the edge device 10, the server device 60 instructs the edge device 10 to perform emergency operation regardless of the model of the equipment 30. The server device 60 can also send instructions to the edge device 10 for the equipment 30 according to the schedule and operations set by the user terminal 70.

[0043] Server device 60 also has the functionality of a web server. The web server responds to requests from client software (web clients) such as a web browser operated by the user and provides the client with screen information written in HTML files, XML, CSS files, JavaScript (registered trademark), etc. An application that uses the web mechanism in this way is called a web application.

[0044] Furthermore, it is preferable that the server device 60 supports cloud computing. Cloud computing refers to a usage model in which network resources are utilized without the user being aware of specific hardware resources. Cloud computing provides users with data and software that they previously used on their own computers, as a service via the network. By providing a web browser that runs on a personal computer or mobile device, and an internet connection environment, users can access a variety of services from any device.

[0045] The user terminal 70 is a client terminal that displays various screens provided by the server device 60. The user terminal 70 may be used by an administrator or by a general user. There are customer-side administrators and management system administrators, but this disclosure does not distinguish between them. Furthermore, the administrator is the person who performs maintenance and management that is not performed by general users who use the equipment 30 on a daily basis.

[0046] The screens displayed by the user terminal 70 can vary widely, but examples include a list screen of the devices 30 and sensor switches 31 connected to the edge device 10 on the customer side, an in-house map showing the locations of the devices 30 and sensor switches 31, and an operation screen for operating the devices 30 and sensors. The screens displayed by the user terminal 70 may also show information about the device 30 and its components that have been found to be malfunctioning.

[0047] The user terminal 70 can be, for example, a PC (Personal Computer), smartphone, tablet, PDA (Personal Digital Assistant), or wearable PC (sunglasses type, wristwatch type, etc.). However, it only needs to have communication capabilities and be able to run a web browser. Alternatively, instead of a web browser, the user terminal 70 may run a dedicated native application for the device management system.

[0048] <Example of placement of learning device and fault detection device> Figure 3 illustrates an example of the arrangement of the fault detection device 400 and the learning device 500 in the equipment management system 100. The learning device 500 is an information processing device that generates a provisional normal prediction model and a normal prediction model during the learning phase of machine learning. The fault detection device 400 is an information processing device that uses the normal prediction model to determine if the air conditioner 32 is faulty.

[0049] In Figure 3(a), the learning device 500 is located on a network connected to the server device 60 and the edge device 10 via network N. The fault detection device 400 may be located on the air conditioner 32, the edge device 10, or the server device 60. Alternatively, the fault detection device 400 may be located independently of these.

[0050] In Figure 3(b), the learning device 500 is integrated with the server device 60. That is, the server device 60 also functions as the learning device 500. The fault detection device 400 may be located in the air conditioner 32, the edge device 10, or the server device 60. The fault detection device 400 may also be located independently of these.

[0051] In Figure 3(c), the learning device 500 is integrated with the edge device 10. That is, the edge device 10 also functions as the learning device 500. The fault detection device 400 may be located on the air conditioner 32, the edge device 10, or the server device 60. The fault detection device 400 may also be located independently of these.

[0052] In Figure 3(d), the air conditioner 32 has a learning device 500 and a fault detection device 400. In other words, the air conditioner 32 serves as both the learning device 500 and the fault detection device 400. The fault detection device 400 may be located independently of the air conditioner 32.

[0053] In all configuration examples, during the learning phase, the learning device 500 acquires learning data (training operation data) from the air conditioner 32 and generates a normal prediction model. During the inference phase, the fault detection device 400 inputs the operating data acquired from the air conditioner 32 into the generated normal prediction model and detects a fault in the air conditioner 32.

[0054] <Hardware configuration of edge device 10 and server device> Next, the hardware configuration of the edge device 10 will be described with reference to Figure 4. Figure 4 is a diagram showing an example of the hardware configuration of the edge device 10. As shown in Figure 4, the edge device 10 has a processor 201, memory 202, auxiliary storage device 203, I / F (Interface) device 204, communication device 205, and drive device 206. Each piece of hardware in the edge device 10 is interconnected via a bus 207.

[0055] The processor 201 has various computing devices such as a CPU (Central Processing Unit). The processor 201 reads various programs into memory 202 and executes them. The processor 201 corresponds to the control unit 110 that controls the entire edge device 10.

[0056] Memory 202 contains main memory devices such as ROM (Read Only Memory) and RAM (Random Access Memory). The processor 201 and memory 202 form a so-called computer, and the processor 201 executes various programs read into memory 202.

[0057] The auxiliary storage device 203 stores various programs and various data used when those programs are executed by the processor 201.

[0058] The I / F device 204 is a connection device that connects the edge device 10 to an example of an external device, such as equipment 30 and sensor switches 31.

[0059] The communication device 205 is a communication device for communicating with the server device 60 via the network N.

[0060] The drive device 206 is a device for setting the recording medium 210. The recording medium 210 here includes media that record information optically, electrically, or magnetically, such as CD-ROMs, flexible disks, and magneto-optical disks. The recording medium 210 may also include semiconductor memory that records information electrically, such as ROMs and flash memory.

[0061] The various programs to be installed on the auxiliary storage device 203 are installed, for example, when the distributed recording medium 210 is set in the drive device 206 and the various programs recorded on the recording medium 210 are read by the drive device 206. Alternatively, the various programs to be installed on the auxiliary storage device 203 may be installed by downloading them from the network N via the communication device 205.

[0062] Figure 5, on the other hand, shows an example of the hardware configuration of the server device 60. Since the hardware configuration of the server device 60 is generally the same as that of the edge device 10, this explanation will focus on the differences between the two. Furthermore, the hardware configuration of the learning device 500, which will be described later, is the same as that of the server device 60.

[0063] The processor 221 reads various programs into memory 222 and executes them. The processor 221 corresponds to the control unit 111 that controls the entire server device 60.

[0064] The I / F device 224 is a connection device that connects the server device 60 to an external device, such as a display device 230 and an operating device 240. The display device 230 displays the internal status of the server device 60. The operating device 240 is used by the administrator of the server device 60 to input various instructions to the server device 60.

[0065] The communication device 225 is a communication device for communicating with the edge device 10 and the user terminal 70 via the network N.

[0066] <About the functions of the learning device> Next, with reference to Figure 6, the functions of the learning device 500 will be described. Figure 6 is a functional block diagram of the learning device 500 according to one embodiment of the present disclosure. The learning device 500 performs two stages of learning. • Stage 1: The learning device 500 generates a hypothetical normal prediction model. • Stage 2: The learning device 500 uses a hypothetical normal prediction model to remove outliers from the training data and generate a normal prediction model.

[0067] Figure 6(a) is a functional block diagram for generating a hypothetical normal prediction model. The learning device 500 includes a learning data acquisition unit 501, a preprocessing unit 502, a first learning data storage unit 503, and a learning unit 504. These functional units of the learning device 500 are functions or means realized by the processor 221 of the learning device 500 executing program instructions loaded into the memory 222.

[0068] The learning data acquisition unit 501 acquires learning data 499. The learning data 499 is normal operating data and has multiple records. Normal operation means that all thermistors, actuators, compressors, and expansion valves of the air conditioner 32 are functioning without abnormalities. The operating data includes explanatory variables and a target variable for generating a provisional normal prediction model. Details of the operating data will be described later, but the operating data of the component whose failure we want to detect is the target variable, and one or more operating data suitable for explaining this target variable are the explanatory variables. Note that the explanatory variables do not include the same operating data as the target variable.

[0069] The preprocessing unit 502 performs preprocessing on the training data. Preprocessing includes, for example, removing transient data and operating data with attached error codes. Transient data refers to operating data acquired during transient states when the room temperature of the air-conditioned space is not in a steady state. Further details will be described later.

[0070] The first learning data storage unit 503 stores the operation data acquired by the learning data acquisition unit 501 and after preprocessing has been completed.

[0071] The learning unit 504 learns from training data using various machine learning algorithms to generate a hypothetical normal prediction model 505. The hypothetical normal prediction model 505 is correspondence information that associates driving data corresponding to explanatory variables with driving data corresponding to the objective variable. In other words, the hypothetical normal prediction model 505 outputs predicted values ​​of driving data corresponding to the objective variable in response to input driving data corresponding to explanatory variables. Such correspondence information can be realized by regression models. Examples of regression models include multiple regression, neural networks, ridge regression, lasso regression, and elastic network regression.

[0072] Note that the driving data includes multiple types of data, and all of them can potentially be the target variable. Therefore, a hypothetical normal prediction model 505 is generated for each target variable.

[0073] Figure 6(b) is a functional block diagram for generating a normal prediction model. The learning device 500 includes a learning data acquisition unit 501, a preprocessing unit 502, a first learning data storage unit 503, an inference unit 506, a valid data identification unit 507, a deviation degree calculation unit 508, a second learning data storage unit 509, and a learning unit 504. Note that the learning data acquisition unit 501, preprocessing unit 502, first learning data storage unit 503, and learning unit 504 may be the same as those in Figure 6(a).

[0074] The inference unit 506 uses the hypothetical normal prediction model 505 to take the driving data corresponding to the explanatory variables from the driving data stored in the first learning data storage unit 503 as input and outputs a predicted value. To distinguish it from the predicted value output by the normal prediction model, this predicted value is called the "hypothetical normal prediction value". The hypothetical normal prediction value is input to the deviation degree calculation unit 508.

[0075] The following data is input to the deviation calculation unit 508. • False normal prediction values ​​output by the false normal prediction model • Measured values ​​stored in the first learning data storage unit 503 (driving data used as the target variable when generating the provisional normal prediction model) The deviation calculation unit 508 calculates the deviation between the measured value (dependent variable) and the assumed normal predicted value. The method for calculating the deviation is as follows, for example. Degree of deviation = Measured value - Predicted normal value Degree of deviation = (Actual value - Predicted hypothetical normal value) / Predicted hypothetical normal value Degree of deviation = Actual value / Predicted normal value By quantifying the degree of deviation between the hypothetical normal predicted value and the actual measured value in this way, we can obtain the difference between the two.

[0076] While it is conceivable to detect outliers using only hypothetical normal prediction values, in this case, outlier detection would depend on the performance of the hypothetical prediction model. Even if the hypothetical prediction model outputs an average value as the hypothetical prediction, if it differs significantly from the actual value, the driving data used to calculate both the actual value and the hypothetical prediction will remain in the training data. Similarly, even if the hypothetical prediction model outputs a hypothetical prediction that deviates significantly from the average, if it is similar to the actual value, the driving data used to calculate both the actual value and the hypothetical prediction will be removed from the training data. In this disclosure, the deviation calculation unit 508 calculates the deviation, thereby suppressing false detection of outliers caused by the performance of the hypothetical prediction model.

[0077] The valid data identification unit 507 identifies valid records from the training data in the first training data storage unit 503 based on statistical information of multiple deviation degrees. In other words, the valid data identification unit 507 identifies valid records by removing outliers from the training data in the first training data storage unit 503. "Valid" means having effectiveness or a positive effect, and in this disclosure, it means identifying operational data that improves the accuracy of the predicted values. Statistical information includes, for example, the mean, median, standard deviation, and quartiles. An example of statistical information is explained in Figures 12 and 13. The valid data identification unit 507 removes operational data (one record including explanatory variables and dependent variables) with a deviation degree above a threshold from the training data in the first training data storage unit 503.

[0078] The valid data identification unit 507 stores the valid training data from the first training data storage unit 503, which has had its valid training data (i.e., training data with a deviation of a threshold or higher) removed, in the second training data storage unit 509. Therefore, the second training data storage unit 509 contains only the training data from which training data estimated to be outliers has been removed.

[0079] The learning unit 504 learns from the valid learning data stored in the second learning data storage unit 509 using various machine learning algorithms to generate a normal prediction model 510.

[0080] <About the functions of the fault detection device> Next, with reference to Figure 7, the functions of the fault detection device 400 will be described. Figure 7 is a functional block diagram of a fault detection device 400 according to one embodiment of the present disclosure. The fault detection device 400 includes an operation data acquisition unit 401, a preprocessing unit 402, an inference unit 403, a deviation degree calculation unit 404, a judgment unit 405, and an output unit 406. These functional units of the fault detection device 400 are functions or means realized by the processor 221 of the fault detection device 400 executing instructions of a program expanded in the memory 222.

[0081] The operation data acquisition unit 401 acquires real-time operation data 410 (current operation data) of the air conditioner 32, one record at a time, which is detected by various sensors (temperature sensor, pressure sensor, etc.) of the air conditioner 32.

[0082] The preprocessing unit 402 performs preprocessing on the operation data 410 acquired by the operation data acquisition unit 401. The content of the preprocessing may be the same as that of the preprocessing unit 502.

[0083] The inference unit 403 uses the preprocessed operating data from the preprocessing unit 402 as input to the normal prediction model 510 to infer predicted values. The input to the normal prediction model 510 is the operating data corresponding to the explanatory variables used to generate the normal prediction model 510, and the predicted values ​​are the predicted values ​​of the operating data corresponding to the objective variable.

[0084] The deviation calculation unit 404 obtains the measured values ​​of the driving data corresponding to the predicted values ​​from the preprocessing unit 402 and calculates the deviation between the predicted values ​​and the measured values. The method for calculating the deviation can be the same as in the learning phase.

[0085] The determination unit 405 compares the deviation degree calculated by the deviation degree calculation unit 404 with a threshold value to determine whether or not the air conditioner 32 has malfunctioned. Since the operating data (actual measured values) in which the deviation degree was large is known, the determination unit 405 can also determine which sensor, actuator, compressor, or expansion valve has malfunctioned.

[0086] The output unit 406 outputs the result of the determination made by the judgment unit 405. For example, the output unit 406 sends an alert to the air conditioner 32, the server device 60, or the user terminal 70. The alert may be sent by displaying a warning on the display of the fault detection device 400 or the user terminal 70, or by notifying the administrator via email or other means that a fault has been detected.

[0087] <Regarding explanatory and dependent variables in training data> Based on Figure 8, we will explain the explanatory and dependent variables in the operating data. Figure 8 shows a list of operating data included in the training data. As an example, let's assume that operating data A to Z are obtained from the air conditioner 32. Operating data A1, A2, ... are operating data A obtained in time series. A set of operating data collected from the air conditioner 32 at the same time is called a record. In Figure 8, operating data A1 to Z1 constitute one record, operating data A2 to Z2 constitute one record, ..., operating data A7 to Z7 constitute one record. In other words, one record contains multiple types of operating data. Specific examples of these will be described later.

[0088] The dependent variable can be any of the driving data A through Z. Furthermore, a single hypothetical normal prediction model 505 or normal prediction model 510 may have two or more dependent variables. The explanatory variables are one or more of the driving data other than the dependent variable. For example, if the dependent variable is driving data A, then the explanatory variables are one or more of the driving data B through Z. All of the driving data B through Z may be used as explanatory variables, or those with a high correlation to the dependent variable, driving data A, may be selected. In this case, driving data A is an example of a dependent variable, and driving data B through Z are examples of explanatory variables. Also, driving data A1 is an example of the first part, and driving data B1 through Z1 are examples of the second part.

[0089] Similarly, if the dependent variable is driving data B, then the independent variables are one or more of the driving data A, C-Z. In this case, driving data B is an example of a second dependent variable, and driving data A, C-Z, etc. are examples of second independent variables (driving data other than the second dependent variable). Furthermore, driving data B1, etc. is an example of a third part, and driving data A1, C1-Z1, etc. are examples of a fourth part. The predictive model that associates driving data B1 with driving data A1, C1-Z1 is an example of a second normal prediction model.

[0090] For example, in the case of multiple regression, the dependent variable and independent variables are associated as follows: y = β0 + β1x1 + β2x2 + ... + β n x n ...(1) y is the dependent variable (e.g., driving data A), x1~x n β0~β are explanatory variables (e.g., driving data B~Z). n These are the coefficients of the explanatory variables. Multiple regression analysis is performed to obtain β0~β n This is determined. The model shown in equation (1) above is the provisional normal prediction model 505 or the normal prediction model 510. As shown in Figure 8, the predicted value PD output by the provisional normal prediction model is calculated for each record in response to the input of explanatory variables. Then, the degree of deviation Δ between this predicted value and the driving data corresponding to the objective variable is calculated for each record.

[0091] Assume that the dependent variable is driving data A, and the independent variables are driving data B, C, and D. If the degree of deviation between the provisional normal prediction values ​​for input to the normal prediction model of driving data B1, C1, and D1 and driving data A1 (actual value) is greater than or equal to a threshold, the effective data identification unit 507 removes at least driving data A1 to D1. Preferably, if the degree of deviation is greater than or equal to a threshold, the effective data identification unit 507 removes one record containing driving data A1 to D1 (driving data A1 to Z1). In this case, the reliability of not only driving data A1 to D1 but also driving data E1 to Z1 may be low. By doing this, the learning device 500 can improve the accuracy of not only the normal prediction model 510 with driving data A as the dependent variable, but also the normal prediction model 510 with driving data B, etc. as the dependent variable.

[0092] As shown in Figure 9, the learning device 500 may generate a hypothetical normal prediction model 505 or a normal prediction model 510 using a neural network. Figure 9 shows the hypothetical normal prediction model 505 or the normal prediction model 510 configured with a neural network 170. The neural network 170 is a form of artificial intelligence (AI) model that trains a computer to process data in a way that mimics the workings of the human brain. An existing configuration will be used for the neural network 170.

[0093] In Figure 9, the inputs to the hypothetical normal prediction model 505 or the normal prediction model 510 are the operating data Ai, Bi, Ci (where i is a natural number), ... and a bias (set to 1). The output is the hypothetical normal prediction value or the predicted value. The hypothetical normal prediction model 505 or the normal prediction model 510 has an input layer 171, a hidden layer 172, and an output layer 173. In the neural network 170 of Figure 9, the L layer is fully connected from the nodes of the input layer 171 to the nodes of the output layer 173. A neural network with deep layers is called a DNN (Deep Neural Network). The layer between the input layer 171 and the output layer 173 is called the hidden layer 172. The number of nodes in the hidden layer 172 is merely an example.

[0094] Weights are assigned to the connections between nodes, and the output from a node multiplied by its weight is passed to the node in the next layer. The node in the next layer receives the outputs of all the nodes in the previous layer, so it sums up the outputs of all the nodes in the previous layer. The node in the next layer activates the summed output with an activation function and passes it to the next node. This process is repeated until the values ​​are passed up to output layer 173.

[0095] In this embodiment, since we want to infer a hypothetical normal prediction value or a predicted value, the hypothetical normal prediction model 505 or the normal prediction model 510 is a regression model (classification models are also available as other models). For this reason, the output layer 173 is provided with one node that outputs a hypothetical normal prediction value or a predicted value.

[0096] Furthermore, the learning phase of the hypothetical normal prediction model 505 or the normal prediction model 510 is assumed to have been learned using existing methods such as backpropagation. In other words, the hypothetical normal prediction model 505 or the normal prediction model 510 learns the correspondence between the driving data Ai, Bi, Ci, etc. and the predicted values.

[0097] In the inference phase using the normal prediction model 510, for example, current operating data A, B, C, etc. are input to the input layer 171. The output layer 173 calculates (infers) the predicted values.

[0098] For each of the driving data A, B, C, etc., extracted from a single record, the output layer 173 outputs a single predicted value. By repeatedly outputting predicted values, time-associated predicted values ​​are obtained.

[0099] The following are specific examples of operating data A to Z in the case of fault detection. ●Examples of explanatory variables are as follows. The type of component indicated in parentheses is a thermistor (temperature sensor), actuator, or other.

[0100] ·Outdoor unit Outdoor temperature (thermistor) Compressor rotation speed (actuator) Fan step (actuator) Main expansion valve opening (actuator) Supercooled expansion valve opening (actuator) Tc target value (other) Te target value (other) ·Indoor unit Suction temperature (thermistor) 1 step (actuator) Indoor expansion valve opening (actuator) Indoor fan tap (actuator) High pressure (other) Low pressure (other) Operation / Stop Information (Other) Remote control temperature settings (other) ●The dependent variable is, for example, the following:

[0101] ·Outdoor unit Discharge pipe temperature (thermistor) Heat exchanger de-icer temperature (thermistor) Supercooled heat exchanger outlet temperature (thermistor) Supercooled heat exchanger liquid tube temperature (thermistor) Heat exchanger liquid tube temperature (thermistor) Accumulator Inlet (Thermistor) Compressor current (other) Compressor rotation speed (actuator) Main expansion valve opening (actuator) Supercooled expansion valve opening (actuator) Fan step (actuator) High pressure (other) Low pressure (other) ·Indoor unit Indoor intake temperature (thermistor) Indoor liquid pipe temperature (thermistor) Indoor gas pipe temperature (thermistor) Indoor fan tap (actuator) Indoor expansion valve opening (actuator) Remote control temperature settings (other) <Outlier Removal Process> Figure 10 illustrates the outlier removal process during the learning phase.

[0102] S1: Driving data Ai, Bi, Ci (where i is a natural number)... are different types of data, each representing one record of driving data. There are at least several hundred to several hundred thousand such records.

[0103] S2: The preprocessing unit 502 of the learning device 500 performs preprocessing on the operation data. Transient data will be explained with reference to Figure 11.

[0104] Figure 11 schematically shows transient data. In Figure 11, the horizontal axis represents the data number assigned to the training data in chronological order, and the vertical axis represents room temperature (one of the operating data). It can be seen that the room temperature increases and then gradually stabilizes. In the case of operating data such as room temperature, the operating data during the process in which the room temperature is changing significantly is transient data. Because transient data is prone to fluctuation, including it in the training data used to generate the provisional normal prediction model 505 or the normal prediction model 510 may reduce the accuracy of the provisional normal prediction value or the normal prediction value. Therefore, the preprocessing unit 502 excludes operating data (1 record) that contains room temperature with a slope 21 above a certain level as transient data. Although temperature is used as an example in Figure 11, rapidly changing operating data can also be excluded as transient data for operating data other than temperature. Note that past operating data is required to calculate the slope 21, and the processing load is high, so the preprocessing unit 502 may consider operating data within a certain time after the compressor starts operating (thermo-on) as transient data.

[0105] S3: Returning to Figure 10, the learning unit 504 generates a hypothetical normal prediction model 505 using the pre-processed training data. Note that multiple hypothetical normal prediction models 505 may be generated from the same driving data by changing the target variable and explanatory variables.

[0106] Furthermore, the operating data collected by the learning device 500 includes information about the source air conditioner 32. This information includes a serial number to identify the individual unit and model number information to identify the model. The provisional normal prediction model 505 and the normal prediction model 510 can be generated from operating data with the same serial number or the same model number information. The provisional normal prediction model 505 and the normal prediction model 510 generated from operating data with the same serial number can accurately detect a failure in that particular air conditioner 32. The provisional normal prediction model 505 and the normal prediction model 510 generated from operating data with the same model number information can accurately detect a failure in an air conditioner 32 of the same model.

[0107] S4: The pseudo-normal prediction model 505 takes the operating data corresponding to the explanatory variables as input when generating the pseudo-normal prediction model 505 and outputs a pseudo-normal prediction value.

[0108] S5: The deviation calculation unit 508 extracts the measured values ​​of the driving data selected as the target variable of the hypothetical normal prediction model 505 from the training data.

[0109] S6: The deviation calculation unit 508 calculates the deviation between the provisional normal predicted value and the measured value.

[0110] S7: The valid data identification unit 507 identifies valid records by removing one record of training data whose deviation is determined to be above a threshold.

[0111] <Method for removing outliers using statistical information> Referring to Figures 12 and 13, we will explain the method for removing outliers using statistical information. Figure 12 shows an example of the distribution of deviation. It is assumed that the deviation follows a normal distribution. When the absolute value of the deviation is large, the difference between the measured value and the assumed normal predicted value is large, so the reliability of the measured value is estimated to be low. Similarly, the reliability of the driving data (explanatory variables) input into the assumed normal predicted value is also estimated to be low. This is because it is unclear whether the abnormality lies in the driving data used for the dependent variable or the driving data used for the explanatory variables.

[0112] For example, if the threshold is set to 2σ to 3σ, the valid data identification unit 507 identifies deviations that do not fall within the mean ± 2σ or are greater than or equal to the mean ± 3σ. The median may be used instead of the mean. The valid data identification unit 507 removes one record from the training data that contains the measured values ​​and the driving data input to the hypothetical normal prediction model, which were used as the basis for calculating the deviation. Therefore, not only the driving data corresponding to the explanatory and dependent variables, but also one record containing the driving data corresponding to the explanatory and dependent variables is excluded.

[0113] Furthermore, as shown in Figure 13, statistical information can also be in the form of quartiles. Refer to Figure 13 to further explain outlier detection using quartiles. Figure 13 is a diagram illustrating the outlier detection method for Q110 using quartiles. Quartile detection is a detection method in which quartiles are found by determining the quartiles (Q1, Q2, Q3) that divide the data equally into four parts when the data is arranged in ascending order, and data that is less than 1.5 times the IQR (interquartile range: Q3-Q1) relative to Q1, or greater than 1.5 times the IQR (Q3-Q1) relative to Q3, is considered an outlier 110. Q1 is called the first quartile, Q2 is called the second quartile, and Q3 is called the third quartile.

[0114] This method can also be considered a way to detect outliers 110 based on statistical information. In the case of quartiles, it is also possible to flexibly adjust whether values ​​greater than or less than N times the IQR are considered outliers 110 by changing N (a real number).

[0115] As shown in Figures 12 and 13, by removing outliers, the valid data identification unit 507 can identify the remaining operating data as valid operating data.

[0116] <<Change the threshold>> The 2σ and 3σ in the normal distribution, or the multiples used to multiply the IQR by 1.5 in the quartiles, do not need to be fixed. This is because the degree of deviation tends to vary when the number of operating data (records) used for training data is small. Therefore, when the number of training data is below a certain level, the valid data identification unit 507 reduces the threshold. This improves the accuracy of the prediction values ​​of the normal prediction model 510 even when the number of training data is small.

[0117] For example, in the case of a normal distribution, if the original threshold is ±3σ, the effective data identification unit 507 changes the threshold to ±1σ or ±2σ if the number of training data is less than a certain amount. Also, for example, in the case of quartiles, if the original threshold is IQR × 1.5, the effective data identification unit 507 changes the threshold to IQR × 1 if the number of training data is less than a certain amount. In this way, the accuracy of the normal prediction model can be improved even with a small amount of training data.

[0118] Alternatively, instead of switching the threshold when the number of training data points falls below a certain level, the threshold may be changed according to the number of training data points.

[0119] Figure 14 shows an example of the correspondence between the number of training data and the threshold. In Figure 14, the horizontal axis represents the number of training data used, and the vertical axis represents the threshold. For example, if the number of training data is 1000 or less, the threshold is ±1σ, and if it is 100000 or more, it is ±3σ. Between 1000 and 100000, the threshold is proportional to the number of training data. In this way, an appropriate threshold can be automatically determined for each number of training data.

[0120] <Action or process> Figure 15 is a flowchart illustrating the process by which the learning device 500 generates a normal prediction model 510. The process in Figure 15 may be performed when a user inputs an operation to start the process. The process in Figure 15 may be performed automatically when enough training data has been collected to generate a normal prediction model 510. The process in Figure 15 may be performed periodically.

[0121] The learning data acquisition unit 501 of the learning device 500 acquires a sufficient number of learning data 499 to generate a normal prediction model, such as one year's worth of operating data (S11).

[0122] Next, the preprocessing unit 502 performs preprocessing on the learning data 499 (S12). The preprocessing unit 502 removes, for example, transient data and operation data to which error codes are attached.

[0123] Next, the learning unit 504 groups together the operating data with the same model number information (S13). Here, a normal prediction model is generated from the operating data with the same model number information, but the learning unit 504 may also group together the operating data with the same serial number. Subsequent processing is performed for each piece of operating data with the same model number information.

[0124] Next, the learning unit 504 extracts driving data corresponding to the objective variable and driving data corresponding to the explanatory variables from the driving data in step S13 (S14). The extraction of driving data corresponding to the objective variable and driving data corresponding to the explanatory variables is performed for each of the hypothetical normal prediction models 505 or normal prediction models 510 to be generated.

[0125] Next, the learning unit 504 generates a hypothetical normal prediction model 505 using a learning method such as multiple regression or a neural network (S15).

[0126] Next, the inference unit 506 inputs the operating data corresponding to the explanatory variables into the hypothetical normal prediction model 505 and calculates the predicted values ​​(S16). If there are multiple hypothetical prediction models 505, the operating data corresponding to the explanatory variables will be different for each hypothetical prediction model 505.

[0127] Next, the deviation calculation unit 508 calculates the deviation between the predicted value and the operating data corresponding to the target variable (S17). If there are multiple hypothetical normal prediction models 505, the operating data corresponding to the target variable will be different for each hypothetical normal prediction model 505.

[0128] Next, the valid data identification unit 507 determines a threshold for determining outliers using statistical information, according to the number of training data (S18).

[0129] Next, the valid data identification unit 507 removes one record containing the driving data corresponding to the objective and explanatory variables used to calculate the deviation from the driving data in step S2 (S19). This allows the valid data identification unit 507 to identify the remaining training data as valid driving data. As an example of removal, the valid data identification unit 507 can ensure that the data is identifiable so that it is not used in generating the normal prediction model 510. Driving data may also be deleted through this removal process.

[0130] Next, the learning unit 504 generates a normal prediction model 510 using the training data from which outliers have been removed, employing learning methods such as multiple regression or neural networks (S20).

[0131] Through the above process, we were able to generate a normal prediction model 510 that accurately calculates predicted values.

[0132] Next, with reference to Figure 16, we will explain the fault detection of the air conditioner 32 using the normal prediction model 510. Figure 16 is a flowchart illustrating the process by which the fault detection device 400 detects a fault in the air conditioner 32.

[0133] The operation data acquisition unit 401 of the fault detection device 400 acquires operation data in real time from the operating air conditioner 32 (S31). Real time means that the maximum delay time is guaranteed.

[0134] Next, the preprocessing unit 402 performs preprocessing on the operation data (S32).

[0135] Next, the inference unit 403 selects a normal prediction model 510 that was generated for the same model number information as the air conditioner 32, which is used to determine whether or not it is faulty (S33).

[0136] Next, the inference unit 403 extracts driving data corresponding to the objective variable and driving data corresponding to the explanatory variable in this normal prediction model 510 (S34).

[0137] Next, the inference unit 403 inputs the driving data corresponding to the explanatory variables into the normal prediction model 510 and calculates the predicted values ​​(S35).

[0138] Next, the deviation calculation unit 404 calculates the deviation between the predicted value and the operating data corresponding to the objective variable (S36).

[0139] Next, the determination unit 405 determines whether the degree of deviation is above a certain level, and if it is above a certain level, it detects a malfunction in the air conditioner 32 (S37).

[0140] Figure 17 is a graph overlaying measured and predicted values ​​over time. The vertical axis represents temperature, and the horizontal axis represents the year the data was acquired. These predicted values ​​are those predicted by the normal prediction model 510. The predicted and measured values ​​almost completely overlap until partway through 2016, confirming the high accuracy of the normal prediction model 510.

[0141] Since mid-2016, the measured values ​​have deviated from the predicted values. On the other hand, the range of fluctuation in the predicted values ​​has remained unchanged since mid-2016, so it is suspected that the sensor detecting the measured values ​​is faulty. For this reason, a service technician repaired it around 2020, and after the repair, the discrepancy between the predicted and measured values ​​almost disappeared.

[0142] <Main effects> In this way, the degree of discrepancy between the predicted value and the actual value by the provisional normal prediction model 505 is calculated, and operating data with a statistically large degree of discrepancy is removed from the training data. By generating the normal prediction model 510 with the training data from which the operating data with a large degree of discrepancy has been removed, failures in the components of the air conditioner 32 can be detected with high accuracy.

[0143] <Reasons why the effect occurs> The first aspect of this disclosure involves "calculating the degree of deviation between the predicted value output by the hypothetical normal prediction model for the input of the second portion of any record and the first portion of the same record as the second portion, for each record in the training data, identifying valid records based on statistical information of multiple deviations, and generating the normal prediction model using the valid records." Since outliers are identified by the hypothetical normal prediction model and the normal prediction model is generated using training data from which outliers have been removed, the accuracy of the predicted values ​​by the normal prediction model can be improved.

[0144] A second aspect of this disclosure involves "removing records containing the first and second portions used to calculate the deviation degree identified as an outlier, and generating the normal prediction model using the training data from which the records have been removed." This allows for the removal of all records that may contain outliers, and since the normal prediction model is generated using the training data from which the records have been removed, the accuracy of the predictions made by the normal prediction model can be improved.

[0145] A third aspect of this disclosure involves "generating the normal prediction model using the training data from which the records have been removed, and generating a second normal prediction model that associates the third part with the fourth part," thus enabling the generation of multiple normal prediction models from a single training dataset by changing the dependent and independent variables.

[0146] A fourth aspect of this disclosure involves "removing records that include the first and second portions used in calculating the deviation, where the deviation is greater than or equal to a threshold from the mean or median," so that outliers can be identified based on the standard deviation when the deviation follows a normal distribution, thus allowing outliers to be identified according to a fixed criterion.

[0147] A fifth aspect of this disclosure involves "removing records that include the first and second portions used in calculating the deviation that are further than a threshold from the first or third quartile," so that outliers can be identified based on quartiles, and thus outliers can be identified according to a fixed criterion.

[0148] A sixth aspect of this disclosure states that "when the number of records included in the training data is less than a certain value, the control unit reduces the threshold value more than when the number of records is not less than a certain value." By reducing the threshold value when the number of records is small, which tends to cause variations in the degree of deviation, the accuracy of the prediction values ​​of the normal prediction model can be improved even with a small number of records.

[0149] The seventh aspect of this disclosure involves "excluding transient data and records with attached error codes from the training data before generating the provisional normal prediction model." Therefore, since the provisional normal prediction model is not created from transient data that is prone to fluctuation and records with attached error codes, the accuracy of the provisional normal prediction model is improved, and the accuracy of identifying deviations that constitute outliers can be improved.

[0150] The eighth aspect of this disclosure involves "inputting into a normal prediction model to calculate a predicted value" and "detecting a failure of the component that output the first portion based on the degree of deviation between the first portion corresponding to the objective variable obtained from the equipment and the predicted value." Thus, a predicted value can be calculated using a normal prediction model, and a failure of a component can be detected based on the degree of deviation between the predicted value and the first operating data.

[0151] The ninth aspect of this disclosure is that "the equipment is an air conditioner," and therefore it can detect a malfunction of the air conditioner. [Explanation of Symbols]

[0152] 30 equipment 32 Air conditioner 60 Server Devices 400 Fault detection device 500 Learning Devices

Claims

1. An information processing device that generates a normal prediction model for predicting the operating data that a device outputs when it is functioning normally, It has a control unit, A hypothetical normal prediction model is generated using training data containing multiple records, with operating data collected simultaneously from the aforementioned equipment treated as one record. The aforementioned operating data is acquired, The aforementioned operating data includes a first part corresponding to a target variable, which is information about temperature or the actuator, and a second part corresponding to an explanatory variable, which is information about temperature or the actuator other than the target variable. The control unit calculates the degree of discrepancy between the predicted value output by the hypothetical normal prediction model for the input of the second portion of any record and the first portion of the same record as the second portion, for each record of the training data. Based on statistical information on multiple deviations, valid records are identified. The normal prediction model is generated using the valid records. Information processing device.

2. The control unit identifies outliers in the deviation degree based on statistical information of multiple deviation degrees, Remove the records containing the first and second portions used to calculate the degree of deviation identified as the outliers. The normal prediction model is generated using the training data from which the aforementioned records have been removed. The information processing apparatus according to claim 1.

3. The operating data collected from the aforementioned equipment at the same time is further comprised of, for each record, a third portion corresponding to a second objective variable which is information relating to temperature or actuators, and a fourth portion corresponding to a second explanatory variable which is information relating to temperature or actuators other than the second objective variable. The control unit generates the normal prediction model using the training data from which the records have been removed, A second normal prediction model is generated that associates the third part and the fourth part. The information processing apparatus according to claim 2.

4. The aforementioned statistical information consists of the mean or median of multiple deviations and the deviation from the mean or median. The control unit removes records that include the first and second portions used to calculate the degree of deviation, where the deviation from the mean or median is greater than or equal to a threshold. The information processing apparatus according to claim 2.

5. The statistical information consists of quartiles of multiple degrees of deviation. The control unit uses N times the IQR (interquartile range) as a threshold, The control unit removes records that include the first and second portions used to calculate the degree of deviation and are located more than a threshold away from the first or third quartile. The information processing apparatus according to claim 2.

6. If the number of records included in the aforementioned training data is less than a certain amount, The control unit makes the threshold smaller than when the number of records is not less than a certain value. The information processing apparatus according to claim 4 or 5.

7. The control unit, From the aforementioned training data, after excluding transient data and records with attached error codes, To generate the aforementioned hypothetical normal prediction model, The information processing apparatus according to claim 1.

8. The control unit, The second portion corresponding to the explanatory variable acquired from the device while the device is operating is input into the normal prediction model to calculate the predicted value. Based on the degree of deviation between the first portion corresponding to the target variable, obtained from the device at the same time as the explanatory variable, and the predicted value, a failure of the component that output the first portion is detected. The information processing apparatus according to claim 1.

9. The aforementioned device is an air conditioner. The information processing apparatus according to claim 1.

10. An information processing device that generates a normal prediction model for predicting the operating data that a device outputs when it is functioning normally, and a model generation method for generating the normal prediction model, The aforementioned information processing device is Control unit and The system includes a hypothetical normal prediction model generated using training data containing multiple records, with each record representing the operating data collected simultaneously from the aforementioned equipment, and the system comprising: The control unit acquires the operation data, The aforementioned operating data includes a first part corresponding to a target variable, which is information about temperature or the actuator, and a second part corresponding to an explanatory variable, which is information about temperature or the actuator other than the target variable. The control unit calculates the degree of discrepancy between the predicted value output by the hypothetical normal prediction model for the input of the second portion of any record and the first portion of the same record as the second portion, for each record of the training data. Based on statistical information on multiple deviations, valid records are identified. The normal prediction model is generated using the valid records. Model generation method.

Images