JP2020170596A - Failure sign detection system - Google Patents

Abstract

To promptly and appropriately determine a sign before a failure of a household system device installed in each home.SOLUTION: A failure sign detection system includes an operating state data acquisition unit 51 that acquires an operating state data group indicating the operating state of a fuel cell cogeneration device 1, a specific feature data setting unit 53 that sets specific feature data that characterizes the specific operating state of the device 1, a machine learning unit 54 trained to input a specific operation state data group indicating the specific operation state of the device 1 and output specific feature data, and a failure sign determination unit 55 that determines the existence or nonexistence of a failure sign on the basis of normal feature data and the specific feature data output by inputting the operation state data group to the machine learning unit 54.SELECTED DRAWING: Figure 3

Description

The present invention relates to a failure sign detection system that estimates a failure sign of a home system device installed in a home.

Patent Document 1 discloses a remote monitoring system that remotely monitors the operating state of a stationary power generation facility and performs preventive maintenance of the power generation facility. This remote monitoring system includes a monitoring terminal that collects operation data of the power generation equipment and a remote monitoring center that diagnoses abnormalities of the power generation equipment based on the operation data transmitted from the monitoring terminal. At the remote monitoring center, the presence or absence of abnormalities and signs of abnormalities in the power generation equipment is determined by determining the upper and lower limits of the analysis results of operating data such as the gas engine speed and cylinder outlet exhaust temperature, and by comparing the results with previous analysis results. .. Furthermore, the administrator, etc. browses the analysis result and the abnormality judgment result on the monitoring personal computer and takes countermeasures.

In the field of semiconductor manufacturing equipment, equipment that employs machine learning for failure prediction is known. In the failure prediction method for semiconductor manufacturing equipment according to Patent Document 2, the feature amount of the device is calculated from the physical quantity indicating the state of the device, and the set of the feature amount in the normal state of the device and the set of the feature amount in the abnormal state of the device are used. The measured feature quantity is evaluated by the machine-learned identification circuit, and if the evaluation result is abnormal, the failure prediction process is performed. Further, for each failure type (abnormal part), the change with time of the feature amount from the normal state to the failure state is registered as the failure model data. At the time of abnormality, the most probable abnormality occurrence site is determined based on the degree of deviation (Mahalanobis distance) between the feature amount vector at each time point of all the registered failure model data and the feature amount vector to be determined. It is also disclosed.

Japanese Unexamined Patent Publication No. 2010-262468 Japanese Unexamined Patent Publication No. 2018-178157

In a remote monitoring system for power generation equipment such as Patent Document 1, since the analysis result of the operation data is simply determined by the upper and lower limits and the comparison with the previous analysis result, the failure sign is accurately estimated. That is difficult. The failure prediction method according to Patent Document 2 may be suitable for failure prediction of a semiconductor manufacturing apparatus installed in a factory and operating under strictly controlled conditions, but is installed in a general household and has various usage conditions. It is not suitable for home system equipment that operates in. In particular, when individually estimating the signs before the failure of the household system equipment installed in each home, the operating state of the household system equipment differs depending on the household, so the application of the failure prediction method according to Patent Document 2 is applicable. Have difficulty.

In view of the above circumstances, there is a demand for a failure sign detection system capable of quickly and appropriately determining a sign before a failure of household system equipment installed in each home occurs.

The failure sign detection system according to the present invention is a failure sign detection system for home system devices installed in each home, and is an operation composed of a plurality of types of time-series operating state data indicating the operating states of the home system devices. An operation state data acquisition unit that acquires a state data group from each household via a data communication network, a specific feature data setting unit that sets specific feature data that characterizes a specific operation state of the household system device, and the household. A machine learning unit learned to input the specific operation state data group indicating the specific operation state of the system device and output the specific feature data, and the normal operation state data group acquired from each household. Is provided with a failure sign determination unit that determines the presence or absence of a failure sign based on the normal time feature data output by inputting the above-learned machine learning unit and the specific feature data.

According to this configuration, the normal time feature data that characterizes the operating state of each household system device is output from the operating state data group indicating the operating state of the household system device acquired from each home via the data communication network. Will be done. Based on the output normal feature data and the specific feature data that characterizes the specific operating state of the home system equipment, for example, these two data are compared (similarity comparison, difference degree comparison, data distance scale comparison). By doing so, it is possible to determine whether or not the household system device to be determined shows a sign of failure. The specific feature data may be a simple numerical value when a plurality of types of operating state data are normalized. In order to determine an accurate failure sign, it is preferable to represent the specific feature data as a vector having one or more elements.

In the first embodiment of the present invention, the machine learning unit inputs the pre-failure operating state data group, which is the operating state data group for a certain period before the failure of the failed home system device, by inputting the pre-failure operating state data group. Learning to output failure feature data as specific feature data, the failure sign determination unit determines the existence or nonexistence of the failure sign by comparing the normal feature data with the failure feature data. In this configuration, among the operation status data groups that are periodically acquired, the operation status data group (fault feature data group) and the failure data that are data before a certain period (for example, one month) before the actual failure occurs. Since the machine learning unit is trained using the time feature data, the failure can be achieved by comparing the normal feature data output by inputting the normal operation state data group with the failure feature data. It is possible to determine the presence or absence of a sign. For example, when the normal feature data and the failure feature data show a predetermined degree of similarity, it can be determined that a failure sign has been detected.

In the second embodiment of the present invention, the machine learning unit inputs the normal operation state data group, which is the operation state data group over a certain period of time in the home system device in the normal state, to obtain the specific feature data. The failure sign determination unit determines the existence or nonexistence of the failure sign by comparing the normal feature data with the normal feature data. In this configuration, since the machine learning unit is trained using the normal operation state data group and the normal time feature data in the home system device in the normal state, it is output by inputting the normal operation state data group. By comparing the normal feature data and the normal feature data, it is possible to determine whether the current home system equipment is in a normal state or an abnormal state. If the normal feature data is more than the permissible range from the normal feature data, the current home system equipment is considered to be in an abnormal state, and a failure sign is determined.

In the third embodiment of the present invention, the machine learning unit inputs the normal operation state data group, which is the operation state data group over a certain period of time in the home system device in the normal state, to obtain the specific feature data. As a result, it is learned to output the normal Mahalanobis distance group, which is the Mahalanobis distance group of the normal operation state data group, and the failure sign determination unit has learned the normal operation state data group acquired from each household. By comparing the normal Mahalanobis distance group output by inputting to the machine learning unit and the normal Mahalanobis distance group, the presence or absence of the failure sign is determined. In this configuration, since the machine learning unit is trained using the normal operating state data group in the normal state home system equipment and the normal Mahalanobis distance group at that time, the normal operating state data group is input. By comparing the normal Mahalanobis distance group and the normal Mahalanobis distance group, it is possible to determine whether the current home system equipment is in a normal state or an abnormal state. If the value of the normal Mahalanobis distance group is more than acceptable with respect to the value of the normal Mahalanobis distance group, the current home system equipment is considered to be in an abnormal state, and a failure sign is determined.

In a preferred embodiment of the present invention, the home system device is a fuel cell cogeneration device. The fuel cell cogeneration system is arranged to supply electric energy and thermal energy to the inside of the house, and is composed of complicated system equipment. Therefore, in the prior art, it is difficult to detect an abnormality (failure) in the operating state and a sign thereof, but by adopting the failure sign detection system according to the present invention, the failure sign detection is improved. In particular, since the hot module, which is a main component of the fuel cell cogeneration apparatus, is expensive, prevention of failure of the hot module greatly contributes to reduction of maintenance cost of the fuel cell cogeneration apparatus.

Other features, actions and effects of the present invention will be clarified by the description of the present invention with reference to the following drawings.

It is a schematic diagram which shows the basic structure of the failure sign detection system. It is a schematic diagram which shows the basic structure of the fuel cell cogeneration apparatus which is an example of a household system equipment. It is a functional block diagram which shows the basic structure of the failure sign detection system. It is a schematic diagram which shows the neural network at the time of learning. It is a schematic diagram which shows the neural network at the time of practice. It is a schematic diagram which shows the structure of the 1st Embodiment of a failure sign detection system. It is a schematic diagram which shows the structure of the 2nd Embodiment of the failure sign detection system. It is a schematic diagram which shows the structure of the 3rd Embodiment of a failure sign detection system.

In the following description of the embodiment, the fuel cell cogeneration apparatus will be taken up as a household system device adopting the failure sign detection system according to the present invention. As shown schematically in FIG. 1, the fuel cell cogeneration apparatus 1 is arranged to supply electric energy and thermal energy to the inside of the house. The failure sign detection system detects a failure sign in the fuel cell cogeneration apparatus. The failure sign detection system is substantially built on the management computer 5 installed in the gas service center. The control unit 2 of the fuel cell cogeneration device 1 and the management computer 5 deployed in each house are connected to each other through a data communication network DN including the Internet and a public line. As a result, various information can be transmitted and received between the management computer 5 and the control unit 2.

As shown in FIG. 2, the fuel cell cogeneration apparatus 1 includes a fuel cell unit 12, a hot water storage tank 11, and a backup heat source machine 13 as components, and extracts hydrogen from city gas and reacts it with oxygen in the air. The heat is generated at the same time to boil water and supply hot water. The fuel cell cogeneration device 1 supplies electric power, hot water supply, and heating to the inside of the house as energy necessary for home life. The control unit 2 controls the fuel cell cogeneration device 1 and generates various data indicating various events such as an operating state in the fuel cell cogeneration device 1.

The data uploaded from the control unit 2 to the management computer 5 includes an operation state data group including a plurality of types of time-series operation state data indicating the operation state of the fuel cell cogeneration device 1. The types of the operating state data are classified according to the states of a large number of sensors and control devices provided in the fuel cell cogeneration apparatus 1 and the amount of change in the states. A time stamp (time) is attached to each data.

Next, the basic principle and configuration of the failure sign detection system will be described with reference to the schematic diagram of FIG. The failure sign detection system built in the management computer 5 includes an operation state data acquisition unit 51, a data storage unit 52, a specific feature data setting unit 53, a machine learning unit 54, and a failure sign determination unit 55.

The operation state data acquisition unit 51 acquires an operation state data group including a plurality of types of time-series operation state data indicating the operation state of the fuel cell cogeneration device 1 sent from each household via the data communication network DN. .. The data storage unit 52 has a database function. The operation state data group acquired by the operation state data acquisition unit 51 is stored in the data storage unit 52 so that it can be extracted by data type, in an arbitrary time band, or both.

The specific feature data setting unit 53 sets specific feature data that characterizes the specific operating state of the fuel cell cogeneration device 1. This specific operating state is predetermined from the time-series state fluctuations in the normal operating state in which the fuel cell cogeneration device 1 is normally operated and the normal operating state, and the time when the fuel cell cogeneration device 1 fails. Includes time-series state fluctuations under the failure sign operation state and the failure sign operation state before time. This specific feature data is data that characterizes the operating state data group acquired in the specific operating state, and can be represented by a vector having one or more elements. By normalizing multiple types of operating state data, it is possible to represent specific feature data with simple numerical values. The specific feature data can also be represented by statistically calculated values estimated from the operating state data group. For example, such statistically calculated values include a correlation matrix, a distance scale, and a Mahalanobis distance in a multivariate space. Etc. are used.

The machine learning unit 54 can be configured here by an optimum model selected from a neural network model such as a recursive neural network model, a recurrent neural network, a convolutional neural network model, and the like as shown in FIGS. 4 and 5. However, recurrent neural networks are advantageous when dealing with time-series data. This neural network model includes an input layer containing seven neurons w1, w2, w3, ..., W7, a first intermediate layer (first hidden layer) containing three neurons x1, x2, x3, and three. It is composed of a second intermediate layer (second hidden layer) containing the three neurons y1, y2, and y3, and an output layer containing the three neurons z1, z2, and z3. In FIG. 4, only the first and second layers are shown as the intermediate layer, but in reality, an input layer corresponding to the amount of data to be input is required, and it is preferable that the hidden layer is also multi-layered. Is.

As shown in FIG. 3, the machine learning unit 54 has a learning unit 541 and a practice unit 542. The machine learning unit 54 learns by the function of the learning unit 541. As shown in FIG. 4, the machine learning unit 54 inputs a specific operation state data group indicating a specific operation state of the fuel cell cogeneration device 1 as learning data, and specifies the specific feature data setting unit 53. Learn to output feature data (teacher data). By incorporating the learning result (learning coefficient) generated by this learning into the practice unit 542, the machine learning unit 54 operates as a learned machine learning unit 54, and as shown in FIG. 5, a normal operating state. When a data group (normal operation state data) is input, normal time feature data is output as output data.

The failure sign determination unit 55 is a normal time feature output by inputting a normal operation state data group into the learned machine learning unit 54, that is, the machine learning unit 54 in which the practice unit 542 operates. The presence or absence of a failure sign is determined based on the data and the specific feature data set by the specific feature data setting unit 53. For example, if the feature data and the specific feature data show a high degree of similarity, it is presumed that the input operation state data group indicates the specific operation state. For example, if the specific operation state is a normal operation state, it can be determined that there is no failure sign at present, and if the specific operation state is a failure operation state (pre-failure operation state), it can be determined that there is a failure sign at present. .. When it is determined that a failure sign exists, the information is transmitted to a maintenance service center or the like, and necessary maintenance service is provided to the corresponding fuel cell cogeneration device 1.

A first embodiment of the failure sign detection system according to the present invention is shown in FIG. In this system, the learning unit 541 of the machine learning unit 54 inputs the operating state data group for a certain period before the failure in the failed fuel cell cogeneration device 1 as learning data, and as specific feature data, the failure feature data. Learn to output.

This failure sign detection system is provided with a preprocessing unit 56 having a pre-failure data extraction unit 561 and a normal data extraction unit 562. The pre-fault data extraction unit 561 extracts the pre-fault operation state data, which is a group of operation state data for a certain period before the failure, from the data storage unit 52 based on the information about the fuel cell cogeneration device 1 in which the failure has occurred. Then, it is given to the learning unit 541 as input data at the time of learning. The specific feature data setting unit 53 gives the failure-time feature data, which is the specific feature data in the first embodiment, to the learning unit 541 as learning-time teacher data. The feature data at the time of failure may be a statistically calculated value of the operation state data group before failure, or may be a vector composed of appropriate elements that characterize the operation state data group before failure.

The machine learning unit 54 operates as a learned unit by giving the practice unit 542 a learning coefficient generated by learning using the pre-failure operation state data and the failure feature data. By inputting the normal operation state data group acquired from each home into the learned machine learning unit 54, the normal time feature data is output.

In this first embodiment, when the input normal operation state data group includes a failure sign, the normal feature data output from the machine learning unit 54 is set by the specific feature data setting unit 53. It will show a high degree of similarity with the characteristic data at the time of failure. When the similarity exceeds a preset threshold value, the failure sign determination unit 55 determines that the fuel cell cogeneration device 1 that has sent the operation state data group is in the state before the failure. When it is determined that a failure sign has occurred, the information is transmitted to a maintenance service center or the like, and necessary maintenance service is provided to the corresponding fuel cell cogeneration device 1.

The operating state data group used in the first embodiment is a main measurement data group which is a basic measurement data group of the fuel cell cogeneration device 1 and a measurement data group in the device constituting the fuel cell cogeneration device 1. It can be divided into sub-measurement data groups. For example, the main measurement data group includes cumulative power generation time, hot module center temperature, stack voltage, stack current, cumulative start count, cumulative shutdown count, reformer inlet temperature, raw fuel gas blower, raw fuel measured value, and hot module outlet. Includes gas temperature, air measurements, etc. The sub-measurement data group includes raw material fuel target value, raw material fuel measurement value, reformed water target value, air target value, heat recovery target temperature, number of reignitions, reforming part inlet temperature, reflux side temperature, exhaust heat recovery water. Includes inlet temperature, exhaust heat exchanger internal temperature, exhaust heat recovery water outlet temperature, auxiliary room upper temperature, tap water temperature, and the like. These measurement data groups are examples, and various other measurement data may be used.

A second embodiment of the failure sign detection system according to the present invention is shown in FIG. In this system, the learning unit 541 of the machine learning unit 54 inputs the normal operating state data, which is a group of operating state data for a certain period of time in the fuel cell cogeneration device 1 in the normal state, as learning data, thereby as specific feature data. Learn to output normal feature data (teacher data).

Therefore, this failure sign detection system is provided with a preprocessing unit 56 having a normal data extraction unit 563 and a normal data extraction unit 562. The normal data extraction unit 563 extracts the normal operation state data, which is the operation state data group during the normal state operation of the fuel cell cogeneration device 1, from the data storage unit 52 and gives it to the learning unit 541 as input data at the time of learning. The specific feature data setting unit 53 gives the learning unit 541 the normal feature data (teacher data) which is the specific feature data in the second embodiment. The normal operation state data may be a statistically calculated value of the normal operation state data group, or may be a vector consisting of appropriate elements that characterize the normal operation state data group.

The machine learning unit 54 operates as a learned unit by giving the learning coefficient 542 generated by learning using the normal operation state data and the normal characteristic data to the practice unit 542. By inputting the normal operation state data group acquired from each home into the learned machine learning unit 54, the normal time feature data is output.

In the second embodiment, when the input normal operation state data group includes a failure sign, the normal feature data output from the machine learning unit 54 is set by the specific feature data setting unit 53. The degree of similarity with the normal characteristic data is low (the degree of difference is high). That is, the operation of the fuel cell cogeneration device 1 is not normal. When the similarity is lower than the preset threshold value, the failure sign determination unit 55 determines that the fuel cell cogeneration device 1 that has sent the operating state data group indicates a failure sign. When it is determined that a failure sign has occurred, the information is transmitted to a maintenance service center or the like, and necessary maintenance service is provided to the corresponding fuel cell cogeneration device 1.

The operating state data group used in the second embodiment may be substantially the same as or different from the operating state data group used in the first embodiment.

A third embodiment of the failure sign detection system according to the present invention is shown in FIG. In this system, the learning unit 541 of the machine learning unit 54 inputs the normal operation state data group, which is the operation state data group over a certain period of time in the fuel cell cogeneration device 1 in the normal state, to perform normal operation as specific feature data. Learn to output the normal Mahalanobis distance group, which is the Mahalanobis distance group of the state data group.

Similar to the second embodiment, this failure sign detection system also includes a preprocessing unit 56 having a normal data extraction unit 563 and a normal data extraction unit 562. The normal data extraction unit 563 extracts the normal operation state data, which is the operation state data group during the normal state operation of the fuel cell cogeneration device 1, from the data storage unit 52, and gives it to the learning unit 541 as input data at the time of learning. The specific feature data setting unit 53 provides the learning unit 541 with the normal Mahalanobis distance group, which is the specific feature data in the third embodiment, as the learning teacher data. The normal Mahalanobis distance group is a Mahalanobis distance calculated using the normal driving state data group, and may be a set of Mahalanobis distances for each type of driving state data (Mahalanobis distance group), or a plurality of Mahalanobis distance groups. It may be the Mahalanobis distance of the Mahalanobis distance group. The Mahalanobis distance group here includes both of them.

The machine learning unit 54 operates as a learned unit by giving the learning coefficient 542 generated by learning using the normal operation state data and the normal characteristic data to the practice unit 542. By inputting the normal operation state data group acquired from each household into the learned machine learning unit 54, the normal Mahalanobis distance group as the normal feature data is output.

In the third embodiment, when the input normal operation state data group includes a failure sign, the normal Mahalanobis distance group output from the machine learning unit 54 is set by the specific feature data setting unit 53. The degree of similarity with the normal Mahalanobis distance group is low (the degree of difference is high). That is, the operation of the fuel cell cogeneration device 1 is not normal. When the similarity is lower than the preset threshold value, the failure sign determination unit 55 determines that the fuel cell cogeneration device 1 that has sent the operating state data group indicates a failure sign. When it is determined that a failure sign has occurred, the information is transmitted to a maintenance service center or the like, and necessary maintenance service is provided to the corresponding fuel cell cogeneration device 1.

Examples of combinations of driving state data groups for calculating the Mahalanobis distance in the third embodiment are listed below;
[Reform water target value, reformed water pump value], [Air target value, air measurement value], [raw fuel target value, raw fuel measurement value], [circuit section input current value, stack current], [1st Combustible gas sensor value, 2nd combustible gas sensor value], [Circuit input current value, Powercon output power value], [Powercon output power value, Stack current], [Powercon output power value, Inverter output current], [Circuit input current Value, inverter output current], [inverter output current, stack current], [raw fuel target value, reformed water target value], [raw fuel measured value, reformed water target value], [tap water temperature, exhaust heat temperature] , [Powercon output power value, hydrogen utilization rate], [Circuit section input current value, hydrogen utilization rate], etc. Of course, measurement data other than these measurement data groups may be used.

[Another Embodiment]
(1) In the above-described embodiment, the trained machine learning unit 54 is provided in the management computer 5. Instead, the trained machine learning unit 54 constructed by the learning coefficient generated by the management computer 5 is a control unit of the fuel cell cogeneration device 1 together with the specific feature data setting unit 53 and the failure sign determination unit 55. You may prepare for 2. In this case, the normal operation state data group generated by the control unit 2 is input to the trained machine learning unit 54 constructed in the control unit 2. It is convenient that the machine learning result data necessary for constructing the learned machine learning unit 54 in the control unit 2 is transferred from the management computer 5 to the control unit 2 using the data communication network DN. , A data transfer method that mediates a smartphone, a portable memory, or the like may be adopted. Further, the control unit 2 in which the machine learning unit 54 constructed by using the machine learning result data is incorporated in advance may be installed in each home. In the method of transferring the machine learning result data via the data communication network DN, the machine learning result data can be periodically generated, and the machine learning unit 54 of the control unit 2 can be constantly updated to the latest version.

(2) The failure sign determination unit 55 can be configured to include not only information on the existence or nonexistence of the failure sign but also the time (predicted time) until the failure occurs. For this purpose, it is preferable to program or table the relationship between the degree of similarity or difference calculated by the failure sign determination unit 55 and the time until failure. At that time, the determination of the failure sign and the estimation of the time until the failure may be performed by separate arithmetic units.

(3) In the description of the first to third embodiments described above, the failure sign detection system is configured to detect the failure sign of the fuel cell cogeneration apparatus. Instead, the failure sign detection system can be configured to detect failure signs of various system devices that manage home life, such as home security system devices, home appliance system devices, cooking system devices, and air conditioning system devices. it can.

The configuration disclosed in the above embodiment (including another embodiment, the same shall apply hereinafter) can be applied in combination with the configuration disclosed in other embodiments as long as there is no contradiction. The embodiments disclosed in the present specification are examples, and the embodiments of the present invention are not limited thereto, and can be appropriately modified without departing from the object of the present invention.

The present invention can be applied to a failure sign detection system of various household system devices such as a fuel cell cogeneration device installed in each home.

1: Fuel cell cogeneration system (household system equipment)
2: Control unit 11: Hot water storage tank 12: Fuel cell unit 13: Heat source machine 5: Management computer 51: Operating state data acquisition unit 52: Data storage unit 53: Specific feature data setting unit 54: Machine learning unit 541: Learning unit 542 : Practical unit 55: Failure sign determination unit 56: Preprocessing unit 561: Pre-failure data extraction unit 562: Normal data extraction unit 563: Normal data extraction unit DN: Data communication network

Claims (6)

It is a failure sign detection system for household system equipment installed in each home.
An operation state data acquisition unit that acquires an operation state data group consisting of a plurality of types of time-series operation state data indicating the operation state of the household system equipment from each household via a data communication network.
A specific feature data setting unit that sets specific feature data that characterizes the specific operating state of the home system device, and
A machine learning unit learned to input a specific operation state data group indicating the specific operation state of the home system device and output the specific feature data.
The presence or absence of a failure sign is determined based on the normal time feature data output by inputting the normal operation state data group acquired from each household into the trained machine learning unit and the specific feature data. Failure sign judgment unit and
Failure sign detection system equipped with. The machine learning unit outputs failure-time feature data as the specific feature data by inputting the pre-failure operation state data group, which is the operation state data group for a certain period before the failure in the failed home system device. Learn to do
The failure sign detection system according to claim 1, wherein the failure sign determination unit determines the existence or nonexistence of the failure sign by comparing the normal feature data with the failure feature data. The machine learning unit inputs the normal operation state data group, which is the operation state data group over a certain period of time in the home system device in the normal state, so as to output the normal feature data as the specific feature data. Learn and
The failure sign detection system according to claim 1, wherein the failure sign determination unit determines the presence or absence of the failure sign by comparing the normal feature data with the normal feature data. The machine learning unit inputs the normal operation state data group, which is the operation state data group over a certain period of time in the home system device in the normal state, to input the Mahalanobis distance of the normal operation state data group as the specific feature data. Learned to output the normal Mahalanobis distance group, which is a group,
The failure sign determination unit includes the normal Mahalanobis distance group and the normal Mahalanobis distance group output by inputting the normal operation state data group acquired from each household into the trained machine learning unit. The failure sign detection system according to claim 1, wherein the presence or absence of the failure sign is determined by comparing the above. The failure sign detection system according to any one of claims 1 to 4, wherein the specific feature data is represented by a vector having one or more elements. The failure sign detection system according to any one of claims 1 to 5, wherein the home system device is a fuel cell cogeneration device.

Images