JP2017192277A - Diagnostic method of power conversion apparatus - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform diagnosis for preventive maintenance by visualizing an operation state of a power conversion apparatus installed, stored, used, etc. in a field.SOLUTION: A diagnostic method of a power conversion apparatus includes: an actual machine learning step S100 including a first neural network configuration data generation step S110 and a first learning step S120 for updating the first neural network configuration data while measuring physical quantity related to a first power conversion apparatus 110; a simulated operation learning step S200 including a second neural network configuration data generation step S210 and a second learning step S220 for updating the second neural network configuration data while measuring physical quantity related to a second power conversion apparatus during simulated operation; and an evaluation diagnosis step S300 including a diagnosis step S320 for diagnosing the operation state of the first power conversion apparatus on the basis of neural network configuration data concerned with a diagnostic template to which second learning is reflected and the first neural network configuration data to which first learning is reflected.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a diagnostic method for a power converter.

2. Description of the Related Art Power converters that perform mutual power conversion between an AC power source and a DC power source, voltage conversion in a DC power source, and the like have recently been represented by a so-called smart grid concept, a HEMS (Home Energy Management System) concept, and the like. In the trend of optimizing the production / supply / consumption of electric power, it has become one of the devices where the needs are increasing.
The power conversion device is connected to a source side that supplies power (for example, one power supply system, a battery, etc.) and a load side that consumes power (for example, other power supply systems, electrical equipment, electrical products, etc.), respectively. Although it is often used, it is not necessarily constant what kind of power supply source and load are connected to these connection destinations. In addition, the environment in which the power conversion device is installed, how it is stored, and how it is operated is not necessarily constant (hereinafter, installation, storage, operation, and the like are collectively referred to as “use”). ).
In this way, once the power converter is placed in the field (use site) that is used after being shipped from the manufacturing plant, the power converter is used because it is used in a variety of ways. Naturally, the “operation status” of how the power conversion device has been operated and operated while it is being performed is also quite different.

Conventionally, a diagnostic method for a power converter using a neural network has been proposed in order to grasp and diagnose the situation or state of the power converter placed in the field (see, for example, Patent Document 1).
In the conventional diagnostic method for a power converter, as shown in FIG. 22, in order to predict the dischargeable time of the secondary battery mounted on the power converter (secondary battery mounted device 930), a simulated operation is performed in advance. The device (secondary battery property evaluation device 910) measures the characteristics of the secondary battery, deepens learning about the relationship between the property and the dischargeable time, and creates neural network configuration data with improved accuracy (see S900). .) In the actual power conversion device (secondary battery mounting device 930), the neural network configuration data created by deepening the learning in advance is taken in and re-deployed as a neural network. (Refer to S910 and S920.) The measured values of the characteristics of the secondary battery measured each time in the actual power conversion device (secondary battery mounting device 930) Are substituted into the input layer of the neural network to obtain a predicted value of dischargeable time of the secondary cell (S930 see.).
In addition, in this specification, the power converter used in an actual field or the apparatus containing it is called "real machine" as a term with respect to the power converter used in a simulation driving | operation apparatus or an apparatus including the same.

  According to the conventional method for diagnosing a power conversion device, the neural network configuration data used on the actual machine side when making a diagnosis is sufficiently learned in advance in the simulated driving device (secondary battery characteristic evaluation device 910) and is accurate. Therefore, there is no need for special learning in the power conversion device (secondary battery mounting device 930) on the actual machine side, and there is no need to accumulate enormous amounts of measurement data. Diagnosis can be made.

JP 2014-206499 A

  However, according to the conventional diagnostic method for the power conversion device, the neural network configuration data used for the diagnosis is not the actual device side (secondary battery mounting device 930) but the simulated driving device (secondary battery characteristic evaluation device). 910) is created by learning uniformly, and based on the neural network configuration data that reflects the circumstances of each actual machine (various usages, operating conditions, etc.) Diagnosis that suits the circumstances cannot be made. In addition, the conventional diagnostic method for a power conversion apparatus diagnoses a “future” situation of predicting a dischargeable time of a secondary battery, and cannot diagnose an operation situation from the past to the present. Furthermore, according to the conventional method for diagnosing a power converter, for each actual machine, according to the past operation status (diagnosis result), maintenance of the power converter is accurately performed, wear, secular change, etc. It is also difficult to accurately carry out “preventive maintenance” such as predicting and carrying out maintenance activities.

  The present invention has been made in view of the above circumstances, and provides a diagnostic method for a power conversion device that can visualize and grasp and diagnose past “operational status” in a real power conversion device. Objective. Furthermore, an object of the present invention is to contribute to accurate maintenance, preventive maintenance, and the like of a power converter by providing a diagnosis result relating to a past operation state for each actual machine.

  The inventor of the present invention uses a neural network to calculate the operation status of the actual power conversion device and the operation status of the power conversion device under a simulated operation environment capable of performing the same operation as the actual device. As a result of earnest development, we were able to know how the actual power converters have behaved and operated in the past (operating conditions) by objectively quantifying and visualizing them and comparing them with each other. The present invention has been completed. The present invention comprises the following elements.

[1] A diagnostic method for a power converter according to the first aspect of the present invention is a diagnostic method for a power converter that uses a neural network to diagnose the operation status of the first power converter. A first neural network configuration data generation step for generating an entity of the first neural network configuration data using a network; and a physical quantity related to the first power converter is measured to obtain a first measurement value; A first learning step including a first learning step of performing first learning by repeatedly updating the first neural network configuration data based on a first measurement value a plurality of times, and using the first neural network A second neural network configuration data generating step for generating an entity of the second neural network configuration data; While performing a simulated operation of the second power conversion device having the same configuration as the force conversion device, a physical quantity related to the second power conversion device is measured to obtain a second measurement value, and based on the second measurement value A second learning step that performs second learning by repeating updating the second neural network configuration data a plurality of times, and the second neural network in which the second learning is reflected As the neural network configuration data related to the diagnostic template, the neural network configuration data related to the diagnostic template and the first neural network configuration data reflecting the first learning are used as the configuration data. And an evaluation diagnosis step including a diagnosis step of diagnosing the operation state.

[2] A diagnostic method for a power converter according to a second aspect of the present invention is a diagnostic method for a power converter that diagnoses the operation status of the first power converter using a neural network. A first neural network configuration data generation step for generating an entity of the first neural network configuration data using a network; and a physical quantity related to the first power converter is measured to obtain a first measurement value; A first learning step including performing a first learning by repeatedly updating the first neural network configuration data based on a first measurement value a plurality of times; (a) the first neural network; A second neural network that generates an entity of second neural network configuration data related to the i-th test mode (i is a natural number) using a network. A network configuration data generation step; and (b) performing a simulated operation in the i-th test mode using a second power converter having the same configuration as the first power converter, and relating to the second power converter. A second learning is performed by repeatedly measuring a physical quantity to obtain a second measured value and updating the second neural network configuration data related to the i-th test mode based on the second measured value a plurality of times. A reference database storage step of storing in the reference database the second neural network configuration data related to the i th test mode in which the second learning is reflected; Execute m times while increasing the number of i until m is an integer of 2 or more, and construct the reference database consisting of m pieces of the second neural network configuration data A reference database construction step, and neural network configuration data corresponding to a test mode for performing an evaluation diagnosis from among the m second neural network configuration data stored in the reference database. A diagnostic template selection step of selecting at least one of the network configuration data, the neural network configuration data related to the diagnostic template, and the first neural network configuration data reflecting the first learning. And a diagnostic step for diagnosing the operation status of the power converter.

[3] In the diagnostic method for the second power conversion device of the present invention, in the evaluation diagnostic step, m stored in the reference database instead of the diagnostic template selection step described in [2]. Among the second neural network configuration data, n pieces of the second neural network configuration data suitable for the diagnosis are selected (n is an integer of 2 or more and n ≦ m), and the n pieces of the second neural network configuration data are selected. A second diagnostic template selection step of selecting at least one of the neural network configuration data corresponding to the test mode for performing the evaluation diagnosis as the neural network configuration data related to the diagnostic template from the two neural network configuration data It is preferable to contain.

[4] In the first and second power conversion device diagnosis methods of the present invention, a hierarchical neural network having an input layer, one or more intermediate layers, and an output layer is used as the neural network. In the first learning step and the second learning step, it is preferable that the first learning and the second learning are performed by an error back propagation method.

[5] In the first and second power conversion device diagnosis methods according to the present invention, in the diagnosis step, a first map is generated based on the first neural network configuration data reflecting the first learning. Creating a second map based on the neural network configuration data related to the diagnostic template, and comparing or collating the pattern of the first map with the pattern of the second map to evaluate the similarity of the pattern, It is preferable to diagnose an operation state of the first power converter according to the similarity.

[6] In the first and second power conversion device diagnosis methods of the present invention, it is preferable that the first power conversion device includes at least an inverter circuit, and at least a battery and a load are connected.

[7] In the first and second power conversion device diagnosis methods of the present invention, in the first neural network, the voltage and current of the battery, the voltage and current of the load, and the temperature of the battery Alternatively, it is preferable that the input layer is constituted by a unit having the temperature of the first power converter as an input factor, and the output layer is constituted by a unit having the discharge time and the charge time of the battery as an output factor.

[8] In the first and second power converter diagnosis methods according to the present invention, the simulated operation of the second power converter is performed under a test mode in which stress is applied to the second power converter. It is preferable that the operation is included.

[9] In the first and second power conversion device diagnosis methods of the present invention, in the actual machine learning step, the first time at a first time when the first learning in the first power conversion device has advanced. Neural network configuration data is acquired and held in association with time information, and further, the first neural network configuration data at a second time after the first time at which the first learning has progressed is acquired. In the evaluation diagnosis step, the neural network configuration data related to the diagnostic template, the first neural network configuration data at the first time, and the second time Diagnosing the operation status of the first power converter based on the first neural network configuration data Preferred.

  According to the first power conversion device diagnosis method of the present invention, a physical quantity related to the first power conversion device (actual machine side) is measured to obtain a first measurement value, and the first neural network is based on the first measurement value. A first learning step in which the first learning is performed by repeatedly updating the network configuration data a plurality of times, and the second neural network configuration data reflecting the second learning by the simulated operation is used as a neural network relating to the diagnosis template As the configuration data, there is a diagnostic step of diagnosing the operation status of the first power conversion device (actual machine) based on the neural network configuration data related to the diagnostic template and the first neural network configuration data reflecting the first learning. According to such a configuration, it is possible to obtain information regarding the past “operation status” in the first power converter of the actual machine as the first neural network configuration data, and based on this, the “operation status” is visualized Can be diagnosed according to the circumstances of each actual machine. In addition, by providing diagnosis results related to the past operation status of the first power converter for each actual machine, it is possible to predict maintenance of the power converter, wear of the power converter, secular change, etc., and carry out maintenance activities. This can contribute to preventive maintenance.

Since the second power conversion device diagnosis method of the present invention basically has the same configuration as the first power conversion device diagnosis method, the same effect as the first power conversion device diagnosis method has. Have
In addition to this, the second power conversion device diagnosis method is described as follows: “Measure the physical quantity related to the first power conversion device,..., Update the first neural network configuration data, the actual machine learning step, the first While performing the mock operation in the i-test mode, the second learning process for performing the second learning is executed m times while increasing the number of i, the reference database construction step for constructing the reference database, and stored in the reference database The neural network configuration data corresponding to the test mode for performing the evaluation diagnosis is selected as at least one of the m second neural network configuration data as the neural network configuration data related to the diagnostic template. Diagnostic template selection process and neural network configuration data related to the diagnostic template And a rating diagnostic steps "include a diagnostic step of first learning to diagnose the operation status of the basis of the first neural network structure data reflected the first power conversion device. According to such a configuration, a large number of second neural network configuration data that are candidates for the neural network configuration data related to the diagnostic template can be prepared as a database under simulated operation in various test modes. From among many possible choices (m second neural network configuration data) stored in the database, the second one that is more similar (matched) to the operating status of the actual power conversion device to be diagnosed 1 Neural network configuration data can be found. By doing so, more accurate and more accurate diagnosis can be performed.

3 is a flowchart for explaining a diagnosis method for the power conversion device according to the first embodiment. It is a block diagram shown in order to demonstrate the diagnostic method of the power converter device which concerns on Embodiment 1. FIG. In the drawings, “neural network” may be simply displayed as “NN” (the same applies to the following drawings). It is a conceptual diagram shown in order to explain the structure of a neural network, and to define a term while showing an example of the 1st neural network in Embodiment 1. FIG. FIG. 3 is a diagram for explaining units constituting the neural network in the first embodiment. FIG. 4A is a conceptual diagram showing one unit UT ^k _j . FIGS. 4B and 4C are graphs showing examples of transfer functions of the transfer unit TF ^k _j constituting the unit. FIG. 4B shows a sigmoid function, and FIG. 4C shows a step function. Is shown. FIG. 3 is a data structure diagram for explaining an example of a data structure of first neural network configuration data and second neural network configuration data in the first embodiment. FIG. 5A is a table relating to the basic architecture of the neural network. FIG. 5B is a model of neural network configuration data for generating the first neural network configuration data and the second neural network configuration data. FIG. 5C is a table relating to the type of transfer function. 6 is a map shown for explaining a change in the state of the neural network when the first learning step S120 of the first embodiment is performed. FIG. 6A is a map representing the state of the neural network as an image based on the first neural network configuration data before starting the first learning. FIG. 6B and FIG. 6C are maps that similarly represent the state of the neural network based on the first neural network configuration data that reflects the first learning. FIG. 6B is a map showing an example of a neural network of a power conversion apparatus that is normally operated under an average environment, and FIG. 6C is a relatively stressed environment under a relatively severe environment. It is a map which shows an example of the neural network of the power converter device in which the driving | running | worked operation was performed. 6 is a map shown for explaining a change in the state of the neural network when the second learning step S220 of the first embodiment is performed. FIG. 7A is a map representing the state of the neural network as an image based on the second neural network configuration data before starting the second learning. FIG. 7B is a map that similarly represents the state of the neural network based on the first neural network configuration data reflecting the second learning. It is a figure shown in order to demonstrate 1st threshold value Sth1 and 2nd threshold value Sth2 used for evaluation of the similarity in diagnostic process S320 of Embodiment 1. FIG. It is a block diagram shown in order to demonstrate the diagnostic method of the power converter device which concerns on Embodiment 2. FIG. It is a figure shown in order to demonstrate the pattern collation of the map in Embodiment 2. FIG. 6 is a flowchart for explaining a diagnosis method for a power conversion device according to a third embodiment. It is a block diagram shown in order to demonstrate the diagnostic method of the power converter device which concerns on Embodiment 3. FIG. It is a data structure figure shown in order to demonstrate an example of the data structure of the reference | standard database in Embodiment 3. FIG. It is another block diagram shown in order to demonstrate the diagnostic method of the power converter device which concerns on Embodiment 3. FIG. It is a block diagram shown in order to demonstrate the diagnostic method of the power converter device which concerns on Embodiment 4. It is a block diagram shown in order to demonstrate the diagnostic method of the power converter device which concerns on Embodiment 5. FIG. It is a block diagram which shows the structure of the system for enforcing the diagnostic method of the power converter device which concerns on Embodiment 6 based on Embodiment 6. FIG. It is a block diagram for demonstrating the 1st power converter device 110 and the real machine 100 in an Example. It is a conceptual diagram for demonstrating the 1st neural network in an Example. It is a conceptual diagram for demonstrating the map which rounded and expressed the 1st neural network in an Example separately. It is a conceptual diagram for demonstrating the 1st map and 2nd map in an Example. FIG. 21A is an example of the first map MAP1 created based on the first neural network configuration data ND1 'in which the first learning in the actual device 100 is reflected. FIG. 21B to FIG. 21D are examples of the second map created based on the second neural network configuration data ND2 ′ reflecting the second learning. It is a figure shown in order to demonstrate the diagnostic method of the conventional power converter device. FIG. 22A is a block diagram of a system related to a conventional method for diagnosing a power converter, and FIGS. 22B and 22C are flowcharts of a conventional method for diagnosing a power converter.

  Hereinafter, a diagnostic method for a power conversion device of the present invention will be described based on an embodiment shown in the drawings. In the drawings, “neural network” may be simply expressed as “NN”, and “database” may be simply expressed as “DB”.

[Embodiment 1]
1. Diagnosis method for power conversion device according to Embodiment 1 The diagnosis method for a power conversion device according to Embodiment 1 can be roughly understood by using a neural network for the operation status of the first power conversion device 110 included in the actual machine 100. This is a diagnostic method for a power conversion device that performs diagnosis, and includes an actual machine learning step S100, a simulated driving learning step S200, and an evaluation diagnosis step S300 (see FIGS. 1 and 2).
The actual machine learning step S100 is performed on the actual apparatus 100 side, and similarly, the simulated driving learning step S200 is performed in the simulated driving apparatus 200. Thereafter, the second neural network configuration data ND2 (described later) reflecting the learning (second learning) in the simulated driving learning step S200 is used as the neural network configuration data NDT (described later) related to the diagnostic template, and the diagnostic template. Diagnosis is performed based on the neural network configuration data NDT and the first neural network configuration data ND1 (described later) reflecting the learning (first learning) in the actual machine learning step S100 (see FIG. 2).

  The “first power conversion device” in the first embodiment is a device included in the actual machine 100, and includes power such as conversion from an AC power source to a DC power source, conversion from a DC power source to an AC power source, and voltage conversion in the DC power source. This is a device having a conversion function. Specifically, it includes units such as converters (AC / DC power supplies), inverters (DC / AC power supplies), DC / DC converters, various rectifiers, etc., and has a predetermined power conversion function, devices, equipment, etc. For example, in addition to single-function devices such as a household storage battery system, an emergency power supply, and a power conditioner, various devices including those functions in part can be used.

  A “neural network” is a model of a brain neural network on a computer (PC, microcomputer, DSP, etc.). In the first embodiment, the “neural network” refers to an ideal existence having a predetermined network structure. And The basic structure of the neural network in the present invention includes a hierarchical structure type, a mutual connection type, and the like without any particular limitation. In the first embodiment, a hierarchical structure type neural network is adopted.

In the “hierarchical neural network”, a plurality of units UT ^k _j simulating brain cell neurons are arranged. These units UT ^k _j belong to any one of the input layer, the intermediate layer, and the output layer. In a hierarchical neural network, one input layer and one output layer are basically provided, but any number of intermediate layers may be provided as long as there are one or more intermediate layers. In this way, the hierarchical neural network is composed of multiple layers, and each layer has a hierarchical structure in which at least one unit UT ^k _j belongs (see FIG. 3).

When the number of layers in the hierarchical neural network, the number of units included in each layer, and the like are determined, the basic shape of the neural network is determined. In the first embodiment, the form of the neural network determined by these pieces of information is referred to as a “neural network basic architecture”.
Each unit UT ^k _j constituting the neural network is connected to each unit in the preceding and following layers by virtual lines such as nerve cell axons (see FIG. 3).
When a signal is input to the neural network, the signal is transmitted in one direction from a unit group belonging to the input layer to a unit group belonging to the output layer via a unit group belonging to the intermediate layer.

Next, the details of the “unit” will be described. When each of the input layer, the intermediate layer, and the output layer is a first layer, a second layer, and a third layer, FIG. 4A shows the kth layer (k = 2, 3,...). The details of the configuration of the j-th (j is a natural number) unit UT ^k _j to which it belongs are shown. The j-th unit UT ^k _j belonging to the k-th layer includes a coupling unit SM ^k _j and a transmission unit TF ^k _j . The combination unit SM ^k _j inputs weighted outputs o ^k−1 _i from all units of the (k−1) _-th layer as shown in the equation (1), and outputs a linear sum x ^k _j thereof. To do.
In Expression (1), the weighting coefficient of the output from the i-th unit of the (k−1) -th layer to the j-th unit of the k-th layer is expressed as W ^{k−1, k} _{i, j} . Further, the output of the _i ^- th unit UT ^k-1 _i of the (k-1) -th layer is denoted as ^ok-1 _i . Transmitting portion ^TF _{k j,} as in the equation (2), and enter the linear combination ^x _{k j} is the output of the coupling unit ^SM _{k j,} through the transfer function _{^{f (x k j), o}} k j = f ( x ^k _j ). As the transfer function f (x ^k _j ), a sigmoid function shown in FIG. 4B, a step function shown in FIG. For example, in the case of a sigmoid function, Expression (2) is used.

  Next, each step constituting the power conversion device diagnosis method according to the first embodiment will be described.

(1) Actual machine learning step S100
The actual machine learning step S100 is performed on the actual machine 100 side, and includes a first neural network configuration data generation step S110 and a first learning step S120 (see FIGS. 1 and 2).

(I) First neural network configuration data generation step S110
First, in the first neural network configuration data generation step S110, the first neural network configuration data ND1 is generated using the first neural network NN1.

The “first neural network NN1” includes, for example, an input layer, two intermediate layers (referred to as a first intermediate layer and a second intermediate layer) and an output layer, and the number of units belonging to the input layer is two, The “basic architecture” is uniquely defined such that the number of units belonging to the first intermediate layer is four, the number of units belonging to the second intermediate layer is four, and the number of units belonging to the output layer is one. An ideal neural network.
The definition relating to the basic architecture of the neural network can also be held as a table TB1 as shown in FIG. For example, the first neural network NN1 in the above example can be defined as having a structure written in a row where the NN number string = 001.

“Neural network configuration data” is data having information necessary for reproducing a neural network on computer software (not only reproducing a framework but also including a learning state). The data structure of the neural network configuration data can take various forms depending on the design policy.
The template representing the data structure of the first neural network configuration data ND1 is defined on software with the basic architecture of the first neural network in mind. In the case of the example shown in FIG. 5B, the weighting coefficient W ^{k− for} each unit UT ^k _j specified by the layer number k and the unit number j in the layer (unit of one row in FIG. 5B). ^1, a variable for storing ^k _{i, j} , a variable for storing the linear sum S _{k, j} , and a variable for storing the type of transfer function are defined.
These variable groups (one row unit) are defined for each unit to which each layer belongs from the first layer to the fourth layer. This template also indirectly includes information about the basic architecture such as the number of layers and the number of units. Note that these variables may be defined (declared) as a structure or may be defined (declared) as an array.

“Generate the substance of the first neural network configuration data ND1 using the first neural network NN1” means a computer (corresponding to the first power converter 110 included in the first power converter 110). The first neural circuit defined (declared) as described above on the software of the first power converter may be a PC, microcomputer, DSP, or the like provided outside the first power converter. Based on the template of the network configuration data ND1, it means to create an entity (instance) so that a value can be stored in a variable.
When an entity as data is generated based on the template defined as described above and a specific value is given to the entity, data called “neural network configuration data” is generated.

For the generated entity, prior to the first learning step, each variable << in the above example, W ^{k-1, k} _{i, j} , ΣW ^{k-1, k} _{i, j} · ok ^-1 _i An initial value may be substituted into (= S _{k, j} ) etc. to establish the initial state of the first neural network configuration data ND1.

  The scale of the first neural network NN1 depends on what is the first power conversion device to be diagnosed, the viewpoint of diagnosis (described later), the ability of the computer to execute the diagnosis method of the power conversion device according to the first embodiment, and the like. Set accordingly.

(Ii) First learning step S120
In the first learning step S120, a physical quantity related to the first power converter 110 is measured to obtain a first measurement value M1, and the first neural network configuration data ND1 is updated based on the first measurement value M1. The first learning is performed by repeating this several times.

First, a physical quantity related to the first power converter 110 is measured to obtain a first measurement value M1.
The “physical quantity related to the first power converter 110” is, for example, the voltage of the battery input / output to / from the first power converter when the power supply source connected to the first power converter 110 is a battery. And physical quantities such as current, load voltage and current, battery BAT temperature or first power converter 110 temperature, battery discharge time and charge time. The “physical quantity related to the first power conversion device 110” in the first embodiment is not limited to the above, but what is / is included in the diagnosis target (first power conversion device) and what kind of viewpoint Can be selected as appropriate depending on whether or not a diagnosis (described later) is performed.
The “first measured value M1” is basically a value obtained by measuring a physical quantity as described above, but is an instantaneous value at a predetermined time (eg, battery voltage at a predetermined time). Alternatively, the value obtained in relation to the time axis such as the discharge time (eg, the time from entering the discharge mode to the end of the discharge mode while charging / discharging is repeated during operation) Alternatively, a value that is fixed from the beginning in terms of specifications, such as battery design specifications (eg, capacity, etc.) may be treated as the first measured value.

Next, the first neural network configuration data ND1 is updated based on the acquired first measurement value M1. That is, based on the first measurement value M1, a so-called “learning” operation is performed, and as a result, the contents of the first neural network configuration data ND1 are updated.
The “learning” may be performed by any method as long as it does not depart from the gist of the invention. However, in the first embodiment, it is suitable for a hierarchical neural network and the calculation method (or algorithm) is compared. The first learning is performed by the error back-propagation method (back propagation method) because the system is simple and systematized and the function approximation capability of the network is theoretically confirmed.
Learning by “error back-propagation method” is explained by applying to the first embodiment, for example, and is acquired while giving a predetermined value (teacher signal) to the output side of each unit belonging to the output layer of the first neural network NN1. A value (learning data) based on the first measurement value is input to a unit belonging to the input layer of the first neural network, and the difference between the output of the unit belonging to the output layer and the teacher signal is equal to or less than a predetermined threshold value. In addition, the weighting coefficients W ^{k−1, k} _{i, j} provided between the units are sequentially adjusted from the output layer side to the input layer side while being adjusted.
When such learning is executed, as a result, values such as the weighting coefficients W ^{k−1, k} _{i, j} and linear sum S _{k, j of} the first neural network configuration data ND1 are updated each time. .

  In the first learning step, such learning that involves updating the first neural network configuration data ND1 is repeatedly performed a plurality of times. As the number of repetitions of learning is increased, the weighting coefficients and the like converge, so that the situation (operation status, etc.) of the actual machine 100 can be reflected deeply, leading to improvement in diagnosis accuracy. On the other hand, the burden on the computer increases as the number of learning increases. Therefore, the number of repeated learnings in one first learning step S120 is appropriately set in consideration of practical aspects.

  The first learning step S120 can be appropriately performed at a timing according to the purpose of diagnosis, intention, neural network specifications, and the like. For example, it can be carried out at the timing when the battery is charged with power in the actual device 100, or when the power is discharged from the battery. By causing the first learning step S120 to be performed on a regular basis, the neural network repeatedly imprints how the power converter 110 has been electrically operated and operated, and the first neural network in which the operation status is reflected in a deep color. -The network configuration data ND1 can be established.

(Iii) For reference, a change in the state of the neural network when the first learning step S120 is performed is shown as an example in FIG.
Here, the circular figure in FIG. 6 represents a unit, and the size of the circle is expressed so as to increase in accordance with the size of the linear sum _{Sk, j} input to the unit. Similarly, the line in FIG. 6 represents a network connecting the units, and the line thickness is expressed so as to increase in accordance with the weighting factors W ^{k−1, k} _{i, j} . In the first embodiment, a state in which the state of the neural network is expressed and visualized as an image based on the neural network configuration data (linear sum, weighting coefficient, etc.) is referred to as a “map”.

A map based on the first neural network configuration data before performing the first learning step S120 is shown in FIG. Before learning, the linear sum in any unit and the weighting coefficient in any network are in the initial value or null state, so here all the circles are circles of the same size, and similarly , Both lines have the same thickness.
Examples of maps based on the first neural network configuration data after performing the first learning step S120 are shown in FIGS. 6B and 6C. For example, according to the first neural network configuration data corresponding to the power conversion device that has been operated normally under an average environment, a map as shown in FIG. 6B is obtained. In addition, according to the first neural network configuration data corresponding to the power conversion device that has been operated under relatively severe conditions under relatively stressful conditions, the map has a shape as shown in FIG. 6C. .
As can be seen from FIG. 6B and FIG. 6C, even if the learning is repeated while measuring the same physical quantity at the same timing, the situation (use of the power converter of the actual machine) It is understood that the different neural network maps established by learning will have different images (in other words, the value of the first neural network configuration data will also be different). it can.

  As described above, by performing (i) the first neural network configuration data generation step S110 and (ii) the first learning step S120, the actual machine learning step S100 is performed, and the learning in the actual apparatus 100 is reflected. Network configuration data ND1 ′ (hereinafter sometimes simply referred to as ND1 ′) can be established.

(2) Simulated driving learning step S200
In the simulated driving learning step S200, a simulated driving environment equivalent to that of the actual machine 100 placed in the field is separately prepared, and simulated driving in various test modes is performed in the simulated driving environment to be learned by the neural network. This is a step of obtaining second neural network configuration data ND2 ′ (hereinafter sometimes simply referred to as ND2 ′), which is important reference information for grasping and diagnosing the operation status of the power converter 110.

  The simulated driving learning step S200 includes a second neural network configuration data generation step S210 and a second learning step S220 (see FIGS. 1 and 2).

(I) Simulated driving device 200
The simulated driving learning step S200 is a step performed in an environment where the second power converter 210 having the same configuration as that of the first power converter 110 can be simulated. For example, the simulated driving learning step S200 is performed in the simulated driver 200. (See FIG. 2).
As the simulated operation apparatus 200, an apparatus equivalent to the actual machine 100 including the first power conversion apparatus 110 may be introduced as the simulated operation apparatus 200 as it is, or even if it is not an actual machine itself, the second power conversion apparatus 210 may be used. Operating environment (temperature, humidity, vibration, electromagnetic wave, etc.) similar to that of the actual machine 100, similar input / output (signal and power input / output), etc. are provided. A device similar to such a jig may be introduced as long as it can simulate the same behavior and operation in the same environment as the conversion device 110.

(Ii) Second neural network configuration data generation step S210
First, in the second neural network configuration data generation step S210, an entity of the second neural network configuration data ND2 is generated using the first neural network NN1.
“Generate the entity of the second neural network configuration data ND2 using the first neural network NN1” is similar to the above on the software of the computer provided corresponding to the second power converter 210. , Generating an entity based on the template of the second neural network configuration data ND2.
This step is basically the same as the first neural network configuration data generation step S110 performed on the actual machine 100 side. However, the point to be noted is that an entity of the second neural network configuration data is generated. Is to use the same neural network (having the same basic architecture) as the first neural network NN1 used on the actual machine 100 side. In the evaluation diagnosis step S300, which will be described later, the first neural network configuration data ND1 and the second neural network configuration data ND2 (the neural network configuration data related to the diagnostic template) can be compared and collated with the same ground. It is for doing so.

(Iii) Second learning step S220
In the next second learning step S220, a second measurement is performed by measuring a physical quantity related to the second power converter 210 while performing a simulated operation of the second power converter 210 having the same configuration as the first power converter 110. Obtaining the value M2 and updating the second neural network configuration data ND2 based on the second measurement value M2 are repeated a plurality of times to perform the second learning.

  The content of the “physical quantity related to the second power conversion device 210”, the way of obtaining the second measurement value M2, the work called “learning”, and the content of the second neural network configuration data ND2 is updated. The learning method and the like have the same contents and conditions as those in the first learning step S120 performed in the actual machine 100.

"Simulation operation" is a test that changes the environment (temperature, humidity, noise, etc.) and operating conditions (applied voltage, current, etc., timing, frequency, number of input / output points, etc.) as appropriate according to the content to be diagnosed. Simulate operation can be performed by setting the mode. For example, a simulated operation can be performed under the following test mode.
(A) Test mode in which stress is applied to the second power converter 210 In the second power converter 210, for example, switching between charging and discharging of the battery is performed more frequently than usual. This is a simulated operation in a test mode in which a relatively stress is applied to a circuit group inside the second power conversion device 210, such as frequent, large power supply noise mixed from a load.
(B) Test mode for promoting aging deterioration This is a simulated operation in a test mode for acceleration of deterioration that promotes deterioration of the circuit group inside the second power converter 210 and elements used therefor. For example, it is possible to set a test mode in which the operating environment temperature of the second power converter 210 is set to a standard high temperature, and the voltage and current are input / output by changing them to the standard width.
(C) Test mode under normal operating conditions Of course, normal operation can also be performed in an average environment under an environment in which the power conversion device is not relatively stressed.

Unlike the first learning step S120 in the actual device 100 installed in the field, the number of learning, the learning frequency, the learning period, and the like are flexible in the second learning step S220 in the simulated driving device 200 according to the contents of the test mode. Can be set. If conditions permit, it is possible to do accelerated learning to deepen learning early and converge the shape of the neural network.
The number of learnings can also be set as appropriate. In general, it is preferable to learn as many times as possible so that the weighting coefficients and the like can be sufficiently converged to a degree that can withstand diagnosis.

(Iv) For reference, FIG. 7 shows an example of a change in the neural network when the second learning step S220 is performed. The notation method is the same as in FIG.
As shown in FIG. 7, by performing the second learning step S220, the neural network (see FIG. 7A) that was initially in the initial state is a neural network whose shape has been changed by reflecting the second learning. A network is formed (see FIG. 7B).

  As described above, by performing (ii) the second neural network configuration data generation step S210 and (iii) the second learning step S220, the simulated driving learning step S200 is performed, and learning in the simulated driving environment (simulated driving device 200). The second neural network configuration data ND2 ′ reflecting the above can be established.

(3) Evaluation diagnosis step S300
As shown in FIGS. 1 and 2, the evaluation diagnosis step S300 is performed in response to the learning result obtained from the actual machine learning step S100 and the learning result obtained from the simulated driving learning step S200. The evaluation diagnosis step S300 includes at least a diagnosis step S320.

(I) Overview of diagnosis step S320
In the diagnostic step S320, the second neural network configuration data ND2 ′ reflecting the second learning is used as the neural network configuration data NDT related to the diagnostic template, and the neural network configuration data NDT related to the diagnostic template and the first learning are Based on the reflected first neural network configuration data ND1 ′, the operation status of the first power converter 110 is diagnosed.

  Here, “the second neural network configuration data ND2 ′ reflecting the second learning as the neural network configuration data NDT related to the diagnosis template” means the second neural network configuration data reflecting the second learning. This means that ND2 ′ is treated as neural network configuration data NDT related to the diagnostic template (hereinafter, the neural network configuration data NDT related to the diagnostic template may be simply referred to as NDT). As a software implementation method, for example, the content (value) of ND2 ′ data may be copied to the NDT entity, or the NDT pointer (ptr) simply indicates the ND2 ′ entity. You may implement | achieve by setting to point. In any case, the content of data extracted as ND2 'is the same as the content of data extracted as NDT.

(Ii) Comparison / collation between neural network configuration data and evaluation of similarity level Next, NDT and ND1 ′ are compared and collated to evaluate the similarity level between neural networks, and the first power conversion device 110 Diagnose the operating status.

  As a method for evaluating the similarity by comparing and collating, the following methods (a) and (b) can be adopted.

(A) Comparison / collation by data value The data value stored in the neural network configuration data NDT related to the diagnostic template and the first neural network configuration data ND1 ′ reflecting the first learning are stored. The degree of similarity can be evaluated by taking out the values of existing data and comparing and collating them. Specifically, there are methods such as the following a) and b).

B) Comparison / collation using thresholds A case where ND1 ′ data and NDT data are compared / collated by focusing on the value of the linear sum S _{k, j} (corresponding to a circle in the map) will be described below. Suppose that the first threshold is Sth1, and the second threshold having a value smaller than the first threshold is Sth2 (see FIG. 8).
The linear sum S _{k, j} related to all the units included in ND1 ′ is compared with the first threshold value Sth1, and only units satisfying S _{k, j} > Sth1 are screened. Record the position information (layer number k, unit number j) of the units that passed the screening. On the other hand, NDT is screened under the same conditions. If the positions of the units that have passed the screening match in ND1 ′ and NDT (the positions of the units that have passed the screening match in ND1 ′ and NDT), they are obtained by learning with an actual machine. It is evaluated that ND1 ′ and the NDT learned by the simulated driving are in the first similarity.
When it is desired to grasp the similarity with higher accuracy, screening is performed under the condition of S _{k, j} > Sth2. Similarly, when the positions of the units that have passed the screening match in each of ND1 ′ and NDT, It can also be evaluated that the ND1 ′ obtained by learning with the actual machine and the NDT learned by the simulation operation are in the second similarity (similarity higher than the first similarity).

B) Comparison / collation focusing on large and small proportions Similarly, the case where comparison / collation focusing on the linear sum _{Sk, j} will be described below. First, in ND1 ′, a unit having the maximum _{Sk, j} value in the same layer is extracted. On the other hand, the unit having the smallest S _{k, j} value is extracted. Record the position information (layer number k, unit number j) of the extracted units. Similar processing is performed in the NDT.
If the extracted maximum unit position matches the minimum unit position in ND1 ′ and NDT, ND1 ′ obtained by learning with the actual machine and NDT learned by the simulation operation are Evaluate that it is in the first similarity.
Although the above is a simple method, in order to further improve the accuracy, the ratio between the maximum S _{k, j} and the minimum S _{k, j} is calculated for each of ND1 ′ and NDT, and the difference between the ratios is calculated. The level of similarity can also be evaluated according to the degree.

In the above a) and b), an example is shown in which comparison / matching is performed on the linear sum S _{k, j} (corresponding to a circle in the map), but the weighting coefficient W ^{k−1, k} _{i, j} (in the map, the line is Can be compared and verified in the same way.
The comparison / collation method and the similarity calculation method are not limited to the above.

(B) Comparison / collation by map A first map MAP1 is created based on the first neural network configuration data ND1 ′ reflecting the first learning, and a second map MAP2 is created based on the neural network configuration data NDT related to the diagnostic template. It is also possible to evaluate similarity by comparing and collating map patterns. Comparison / collation by this method will be described in detail in the section [Embodiment 2].

(Iii) Diagnosis of the operation state of the first power conversion device according to the similarity degree As described above, after completing the comparison / collation and evaluation of the similarity degree of both neural networks based on ND1 ′ and NDT, the diagnosis step In S320, the operation status of the first power converter 110 is diagnosed according to the similarity.

  Examples of the “operation status” to be diagnosed include the following viewpoints. In the past, what kind of input and output were performed in what environment and what kind of stress was applied to the first power converter 110. In the past, which part of the circuit group constituting the first power conversion device 110 is mainly subjected to stress and the deterioration of the elements and the like has progressed.

The diagnosis of the operation status of the first power converter according to the degree of similarity in the diagnosis step S320 is performed as follows, for example.
First, the degree of similarity between the first neural network configuration data ND1 ′ and the neural network configuration data NDT related to the diagnostic template (taken from ND2 ′ and treated as NDT) is grasped.
If the degree of similarity is slightly high, it is estimated that the first power converter 110 may have been operating in an environment / operating condition similar to the test mode in which the simulated operation is performed.
In addition, when the similarity is higher than the above, or when it can be determined that the similarity is extremely high and the shapes of the two neural networks are substantially the same, the first power converter 110 performs the simulation operation. It is estimated with high probability that it has been operating under the same environment and operating conditions as the test mode.
As described above, the second neural network configuration data ND2 ′ in various test modes is extracted, and the similarity with the first neural network configuration data ND1 ′ on the actual machine 100 side is evaluated as described above. The test mode corresponding to the second neural network configuration data ND2 ′ having the highest similarity (by simulated operation) is specified. When the first power conversion device 110 is identified in this way, it can be estimated with a considerable degree of accuracy that the actual first power conversion device 110 has been operated under the environment and operating conditions equivalent to the identified test mode.
In this way, it is possible to visualize and grasp the “operation status” of the first power converter 110 and perform a diagnosis in accordance with the actual situation of each actual machine.

Further, for example, after preparing the second neural network configuration data ND2 (which will be treated as NDT later) according to the “(b) test mode for promoting aging degradation” in the simulated driving learning step S200, the actual machine 100 side When the first neural network configuration data ND1 ′ is compared with the NDT and it is determined that the degree of similarity is higher than a predetermined level, the aging of related circuits, elements, etc. in the first power converter 110 It can also be seen as a sign that deterioration is progressing. When such a diagnosis result is given, preventive maintenance such as inspection and replacement centering on the corresponding circuit, element, etc. of the first power converter 110 can be performed at the next actual machine maintenance.
Thus, according to the diagnostic method for the power conversion device according to the first embodiment, by providing a diagnosis result related to the past operation status of the first power conversion device for each actual machine, accurate maintenance of the power conversion device and In addition, it is possible to contribute to preventive maintenance and the like in which maintenance activities are performed by predicting wear of the power conversion device, secular change, and the like.

2. Effect of power conversion device diagnosis method according to Embodiment 1 (1) According to the power conversion device diagnosis method according to Embodiment 1, a physical quantity related to the first power conversion device 110 (actual machine side) is measured and The first learning step S120 for obtaining the first measurement value M1, and updating the first neural network configuration data ND1 based on the first measurement value M1 a plurality of times to perform the first learning, and the simulation operation The second neural network configuration data ND2 ′ reflecting the second learning is set as the neural network configuration data NDT related to the diagnostic template, and the first neural network configuration data NDT related to the diagnostic template and the first learning are reflected. Diagnostician for diagnosing the operation status of the first power converter 110 based on the neural network configuration data ND1 ′ With the S320. According to such a configuration, it is possible to obtain information regarding the past “operation status” in the first power converter of the actual machine as the first neural network configuration data, and based on this, the “operation status” is visualized Can be diagnosed according to the circumstances of each actual machine. In addition, by providing diagnosis results related to the past operation status of the first power converter for each actual machine, it is possible to predict maintenance of the power converter, wear of the power converter, secular change, etc., and carry out maintenance activities. This can contribute to preventive maintenance.

(2) In the diagnostic method for the power conversion device according to the first embodiment, a hierarchical neural network having an input layer, one or more intermediate layers, and an output layer is used as the neural network. In the learning step and the second learning step, the first learning and the second learning are performed by the error back propagation method.
In this way, since the neural network of the hierarchical structure type is used and learning by the error back propagation method is performed, rather than the diagnostic method of the power conversion device based on other types of neural networks such as a mutual coupling type, Since the size of the neural network configuration data can be made relatively small, and the computation can be made relatively light. As a result, a large amount of CPU resources are not required and learning can be performed in a short time. A diagnostic method for a power converter that is commercially / practically advantageous can be provided.

[Embodiment 2]
1. Diagnosis method for power conversion device according to embodiment 2
The diagnostic method for the power converter according to the second embodiment basically has the same configuration as the diagnostic method for the power converter according to the first embodiment. However, in the diagnostic step S320, the neural network configuration data is mapped and diagnosed. Is different from the diagnosis method of the power conversion device according to the first embodiment in that
In other words, as shown in FIG. 9, the diagnosis method for the power conversion device according to the first embodiment is the first based on the first neural network configuration data ND1 ′ in which the first learning is reflected in the diagnosis step S320. A map MAP1 is created, a second map MAP2 is created based on the neural network configuration data NDT related to the diagnostic template, and the pattern similarity is compared or collated with the pattern of the first map MAP1 and the pattern of the second map MAP2. And the operation status of the first power converter 110 is diagnosed according to the similarity.

(1) Map expression method A “map” is a visual representation of the state of a neural network as an image in some form, whether it is a planar expression or a three-dimensional expression. I do not care. In the following, the description will be continued on the premise of a map expressed in a planar manner.
For example, FIGS. 6 and 7 mentioned above are maps. The circle represents the unit UT ^k _j , the line connecting the circles represents a network connecting the units, the size of the circle represents the linear sum S _{k, j} input to the unit, and the thickness of the line represents the weighting coefficient W ^{k−1. , K} _{i, j} respectively. In addition to using different sizes / thicknesses, expressions such as different colors and different types of lines (dotted lines, broken lines, solid lines, double lines, etc.) are used alone or in combination. The map can also be expressed.

(2) Diagnosis Using Map (i) Creation of First Map MAP1 First, the first map MAP1 is created based on the first neural network configuration data ND1 ′ reflecting the first learning. Taking the case where the unit UT ^k _j is arranged as an example, the size of the circle is determined according to the value of the data (S _{k, j} ) of ND1 ′, and the position of the unit is determined in advance. (Here, k-th layer, j-th position).
Similarly, when arranging the line corresponding to the network, the thickness of the line as a figure is determined according to the size of the data of ND1 ′ (weighting coefficient W ^{k−1, k} _{i, j} ), The line is placed at a predetermined network position.
When the circles and lines are arranged corresponding to all the units and networks, the first map MAP1 is completed.

(Ii) Creation of the second map MAP2 The second map MAP2 is created based on the neural network configuration data NDT related to the diagnostic template, in the same manner as the first map creation procedure described above.

(Iii) Comparison / collation between first map MAP1 and second map MAP2 When ND1 ′ and NDT are visualized and first map MAP1 and second map MAP2 are created, both maps are compared and collated. We examine whether a similar tendency is seen in the shape of the map and evaluate the similarity.

As a comparison / collation method, various methods including the following (a) to (c) can be adopted.
(A) A comparison / collation between the first map MAP1 and the second map MAP2 is performed using an image processing technique (a pattern matching generally used in a production line or a recognition technique used in face authentication or the like). It can be carried out.
For example, as shown in FIG. 10, first, an image of the second map MAP2 (neural network configuration data related to the diagnostic template) is captured by the camera, and binarization processing is performed on the captured image data (for example, FIG. 10 ( A binarization process such as a) is performed.) On the other hand, the same binarization process is performed on the first map MAP1 (see FIG. 10B). Pattern matching processing is performed by collating both images subjected to the binarization processing.
If both images match, both maps are evaluated to be “very high similarity” (substantially the same). Even if the two images do not completely match, it may be determined that they are almost the same by reducing the severity of collation. In this case, it is evaluated as “relatively high similarity”. In the above description, the example in which the similarity is evaluated by the pattern matching process of the binarized image has been described. However, the present invention is not limited to this. For example, feature points as indicated by x in FIG. 10 can be extracted and the similarity between both maps can be evaluated based on the feature point information.

(B) With regard to the size of the circle, both maps can be compared and collated by paying attention to a threshold value, a proportion of magnitude, and the like.
A circular area is grasped from the camera, and based on the value of the area [1. (3) Similar to the method described in (ii) (a) Comparison / collation by data value], a method using a threshold (whether the area of the circle exceeds a predetermined threshold as in the above method) The degree of similarity can be evaluated by a method (determination by extracting the circle with the largest area and the circle with the smallest area) focusing on the proportions of large and small (see FIG. 8 for details). See section).

(C) The first map MAP1 and the second map MAP2 can be compared and verified by human eyes. As an example, first, a bird's-eye view of both maps is used, and attention is paid to the circle to grasp the characteristics of the distribution of the size of the circle. Determine whether there is a common point in the position of the unit where the large and small features appear, such as the position of the largest circle, the smallest circle, etc., and if there are many common points, evaluate that the similarity of both maps is high be able to. Furthermore, as mentioned in the above method, the degree of similarity should also be evaluated based on whether the ratio of the size of the large circle unit to the small circle unit is close in both maps (ratio of the ratio). Can do.

  In the above description, the unit size (circular size representing the size of the value of the linear sum) has been described with reference to FIG. 10, but the thickness of the network (the value of the weighting coefficient value is large). Comparison / collation can be performed in the same manner for the thickness of the line representing the thickness. Also, the method of comparing and collating the magnitude of the value of the linear sum using a map and comparing and collating the magnitude of the value of the weighting coefficient directly using the data value is used. Also good.

(Iv) Evaluation and Diagnosis by MAP1 Single Unit In the above, an example of evaluation by both maps (MAP1 and MAP2) has been described. However, it is also possible to evaluate the operation state of the actual machine 100 using only the map MAP1 single unit based on ND1 ′. it can.
For example, in FIG. 6 (c), the which second unit UT ¹ ₂ of the circular first layer (input layer) is increased, frequently there is an input by the relatively large value for this second unit UT ¹ ₂ It can be estimated that the physical quantity corresponding to this input is often a relatively large value. Further, the line is thick between UT ¹ _{2 to} UT ² ₃ and between UT ¹ _{1 to} UT ² ₂ , and during this time, it can be understood that the signal is transmitted relatively frequently. It can be estimated that the relationship between UT ¹ _{2 to} UT ² ₃ and that between UT ¹ _{1 to} UT ² ₂ are high.

2. Effects of the power conversion device diagnosis method according to the second embodiment (1) In the power conversion device diagnosis method according to the second embodiment, the first is based on the first neural network configuration data ND1 ′ reflecting the first learning. A map MAP1 is created, and a second map MAP2 is created based on the neural network configuration data NDT related to the diagnostic template. In this way, the state of the neural network obtained by accumulating learning on the actual actual machine 100 is expressed as the past “operational situation (how has it behaved and operated) of the first power converter 110. ”Is visualized as“ map ”. Therefore, by looking at the first map MAP1, it is possible to grasp the “operation status” as an image.

(2) Moreover, in the diagnostic method of the power converter device which concerns on Embodiment 2, the pattern similarity is evaluated by comparing or collating the pattern of a 1st map with the pattern of a 2nd map, and according to this similarity The operating status of the first power converter is diagnosed.
Since the patterns of both maps visualized in this way are compared and collated with each other, not only a diagnosis based on binary evaluation such as matching / mismatching of maps but also an intermediate value between two values of “similarity” of patterns. It is possible to evaluate a region (a similar relationship located in the middle between both maps having the same pattern and not similar to each other). Therefore, instead of a binary assertive diagnosis, the actual machine 100 includes a “probability or high possibility” such as “possibility” of operating in the same environment and operating conditions as the test mode. Diagnosis can be made based on the information.

(3) Further, according to the diagnostic method for the power conversion device according to the second embodiment, as described above, visualization is performed as a “map”, so that the diagnostic step S320 can be automated by an image processing technique such as pattern matching. . It is also relatively easy for a human to visually perform part of the diagnosis step S320 (as if a doctor grasps a plurality of MRT images visually and diagnoses the “trend or likelihood of disease” of the disease. as.).

  The power conversion device diagnosis method according to the second embodiment has the same configuration as the power conversion device diagnosis method according to the first embodiment except for the configuration described above, and thus the power conversion device diagnosis method according to the first embodiment. Among the effects possessed, the corresponding effect is maintained as it is.

[Embodiment 3]
1. Diagnostic method for power conversion device according to embodiment 3
The power conversion device diagnosis method according to the third embodiment basically has the same configuration as the power conversion device diagnosis method according to the first embodiment, but has a reference database construction step S400 and is stored in the reference database. A diagnostic method for the power conversion device according to the first embodiment in that the second neural network configuration data to be diagnosed is selected from the m second neural network configuration data that are to be diagnosed and attached to the diagnosis; Is different.
That is, as shown in FIGS. 11 and 12, the diagnostic method for the power conversion device according to the third embodiment includes (a) the i-th test mode (1) using the first neural network NN1 in the reference database construction step S400. a second neural network configuration data generation step S410 for generating an entity of the second neural network configuration data ND2 relating to i being a natural number), and (b) a second power conversion device having the same configuration as the first power conversion device 110 While performing a simulated operation in the i-th test mode using 210, a physical quantity related to the second power converter 210 is measured to obtain a second measured value M2, and the i-th test is performed based on the second measured value M2. A second learning step S42 for performing second learning by repeatedly updating the second neural network configuration data related to the mode a plurality of times. When, and are (c) and the reference database storing step and one step block for storing second neural network structure data ND2 'to the reference database according to the second learning is reflected first i test mode. Then, with respect to these one process block, i is executed m times while increasing the number of i from 1 to m (m is an integer of 2 or more), and a reference database consisting of m second neural network configuration data ( In FIG. 12, the reference DB is represented.).
Similarly, as shown in FIGS. 11 and 12, in the evaluation diagnosis step S300, m pieces of second neural network configuration data ND2 ′ stored in the reference database (indicated as reference DB in FIG. 12) are stored. A diagnostic template selection step S310 is performed in which at least one of the neural network configuration data corresponding to the test mode for performing the evaluation diagnosis is selected as the neural network configuration data NDT related to the diagnostic template. Similar to the diagnostic method of the power conversion device according to the first embodiment, the first power conversion device 110 is based on the neural network configuration data NDT related to the diagnostic template and the first neural network configuration data ND1 ′ reflecting the first learning. Diagnosis process to diagnose the operating status of Has become shall implement 320.

In the diagnostic method for the power conversion device according to the third embodiment, m simulations are performed under conditions << i-th test mode (i is an integer of 2 or more from 1 to m) >> according to various variations. The second neural network configuration data ND2 ′ is prepared.
As a variation of the test mode, [1. (2) (iii) As mentioned in the section of the second learning step S220], (a) a test mode in which stress is applied to the second power converter 210, (b) a test mode in which aging degradation is promoted, ( c) The test mode under normal operating conditions can be set as appropriate.

The reference database (reference DB) may have any configuration as long as it can store m types of second neural network configuration data ND2 ′, appropriately select (select) them, and individually extract the data. For example, the second neural network configuration data ND2 ′ _i can be individually stored as records, and a data structure in which these records are managed as a list by a pointer (ptr) can be adopted (see FIG. 13). .

2. Effects of the power conversion device diagnosis method according to the third embodiment The power conversion device diagnosis method according to the third embodiment basically has the same configuration as the power conversion device diagnosis method according to the first embodiment. 1 has the same effect as that of the power conversion device diagnosis method according to 1.
In addition to this, according to the diagnostic method of the power conversion device according to the third embodiment, “the physical quantity related to the first power conversion device is measured, and the first neural network configuration data is updated. While performing a simulation operation in the learning step, the i-th test mode .... a second learning step in which the second learning is performed, ... is executed m times while increasing the number of i, and a reference database is constructed to construct a reference database. At least one of the m second neural network configuration data stored in the step and the reference database is used as the neural network configuration data related to the diagnostic template. A diagnostic template selection process for selecting one of them, and a neural template related to the diagnostic template And a diagnostic step of diagnosing the operating status of the first power converter based on the network configuration data and the first neural network configuration data reflecting the first learning. Thus, a large number of second neural network configuration data that are candidates for the neural network configuration data related to the diagnostic template can be prepared as a database. From among many possible choices (m second neural network configuration data) stored in the database, the second one that is more similar (matched) to the operating status of the actual power conversion device to be diagnosed 1 Neural network configuration data can be found. By doing so, more accurate and more accurate diagnosis can be performed.

As shown in FIG. 14, m pieces of second neural network configuration data ND2 _i are respectively set as neural network configuration data NDT _i related to the diagnostic templates, and the neural network configuration data NDT _i related to the respective diagnostic templates are Based on the first neural network configuration data ND1 ′, the degree of similarity of the neural network based on the neural network configuration data NDT _i and the first neural network configuration data ND1 ′ related to each diagnostic plate is evaluated. The evaluation diagnosis may be comprehensively performed with reference to the evaluated m similarities.

  12 and FIG. 14 do not describe how diagnosis is performed using maps, but in the third embodiment as well, the first neural network configuration data ND1 ′ based on the first learning reflects the first learning. The diagnosis using the first map MAP1 and the second map MAP2 based on the neural network configuration data NDT related to the diagnosis template can be similarly applied. In this case, the power conversion device diagnosis method according to the second embodiment has the same effect.

[Embodiment 4]
The diagnosis method for the power conversion device according to the fourth embodiment basically has the same configuration as the diagnosis method for the power conversion device according to the third embodiment, except that m second neural nodes stored in the reference database are included. It differs from the diagnosis method of the power converter according to the third embodiment in that the appropriate second neural network configuration data is further narrowed down to n from the network configuration data.
That is, as shown in FIG. 15, the diagnostic method for the power conversion device according to the fourth embodiment is stored in the reference database instead of the diagnostic template selection step S310 described in the third embodiment in the evaluation diagnostic step S300. N pieces of second neural network configuration data suitable for the diagnosis are selected from the m second neural network configuration data ND2 ′ (n is an integer of 2 or more and n ≦ m), and the n The second neural network configuration data is selected from at least one of the second neural network configuration data as neural network configuration data NDT related to the diagnostic template, corresponding to the test mode for performing the evaluation diagnosis. The diagnostic template selection process S315 (not shown) is included.

For example, if it is known that the first power converter 110 tends to be used mainly in a specific season, m pieces of second neural network configuration data ND2 ′ stored in the reference database Of these, n neural network configuration data groups suitable for the season to be diagnosed can be selected before being subjected to the diagnosis step S320.
In the reference database construction step S400, second neural network configuration data groups for summer, winter, and spring / autumn are prepared in advance, respectively, and in the second diagnostic template selection step S315, “season” to be diagnosed. Only n second neural network configuration data groups that match the above are selected. By doing so, more efficient diagnosis can be performed than when comparing and collating all m patterns including data of all seasons. Further, since the second neural network configuration data suitable for the season related to diagnosis is selected and narrowed down, more accurate diagnosis can be performed.

Furthermore, similarly, when it is known that the load connected to the first power converter 110 is a specific electrical facility or electrical product (such as a heating appliance), the type of load (above In the example, n second neural network configuration data groups corresponding to the heating appliances) can be selected and diagnosed.
In this way, in the above example, it is possible to perform a more efficient diagnosis than performing a diagnosis by hitting all m patterns for all types of loads. In the above example, the situation (operation status) of the first power converter 110 differs depending on the type of load to be connected. Therefore, according to the diagnosis method for the power converter according to the fourth embodiment, the diagnosis is related to the diagnosis. More accurate diagnosis can be performed by selecting and narrowing down the second neural network configuration data suitable for the device.

In the above description, an example in which the second neural network configuration data group is selected from the viewpoint of when the “season” to be diagnosed has been described. However, the actual neural network 100 also has the first neural network configuration for each season. Data ND1 is generated and held, and diagnosis is performed based on the first neural network configuration data corresponding to the season to be diagnosed and n second neural network configuration data groups corresponding to the season. It may be.
In the diagnosis according to the type of load in the above example, it is assumed that the second neural network configuration data group corresponding to different types of load is included in one reference database. The reference database may be constructed separately for different types of loads.

  According to the diagnostic method for the power conversion device according to the fourth embodiment, the number m of second neural network configuration data is selected as the neural network configuration data related to the diagnostic template, and the number of diagnostic templates is narrowed down. , It is possible to perform a more efficient diagnosis than performing a diagnosis for all m patterns. Further, since the second neural network configuration data suitable for diagnosis is selected and narrowed down, more accurate diagnosis can be performed.

  The power conversion device diagnosis method according to the fourth embodiment has the same configuration as the power conversion device diagnosis method according to the third embodiment except for the configuration described above, and thus the power conversion device diagnosis according to the third embodiment. It has the corresponding effect as it is among the effects of the method.

[Embodiment 5]
The diagnosis method for the power conversion device according to the fifth embodiment basically has the same configuration as the diagnosis method for the power conversion device according to the first embodiment, but the first neural network at different times in the same real machine 100. It differs from the diagnostic method of the power converter according to the first embodiment in that the configuration data (ND11 ′ and ND12 ′) is acquired and diagnosed.
That is, in the diagnosis method for the power conversion device according to the fifth embodiment, as shown in FIG. 16, in the real machine learning step S100, the first neural network at the first time t1 when the first learning in the real machine 100 has advanced. The configuration data ND11 ′ is acquired and held while being associated with the time information, and the first neural network configuration data ND12 ′ at the second time t2 after the first time t1 in which the first learning in the real machine 100 further proceeds is obtained. In the evaluation diagnosis step S300, the neural network configuration data NDT related to the diagnostic template, the first neural network configuration data ND11 ′ at the first time, and the second time are acquired and stored in association with the time information. Based on the first neural network configuration data ND12 ′ in FIG. The operation state of the power converter 110 is diagnosed.

The diagnosis method for the power conversion device according to the fifth embodiment is particularly useful when the first power conversion device 110 is regularly diagnosed such as regular maintenance of the actual device 100.
For example, the respective states (forms) of the first neural network configuration data ND11 ′ at the first time t1 and the first neural network configuration data ND12 ′ at the second time t2 are compared, and ND12 is compared with ND11 ′. If 'is significantly changed, it is assumed that some variation has occurred, and depending on the changed shape of the neural network, the changed location (unit, network), and the degree of change From the point of view, you can perform more detailed maintenance.

When the state (form) of the first neural network configuration data ND12 ′ at the second time t2 can be recognized as similar to the state (form) of any second neural network configuration data ND2 ′. Can compare the three neural network configuration data of ND11 ′, ND12 ′, and ND2 ′ with each other to estimate the degree of progress toward the secular change, and to capture their signs.
That is, when the similarity between ND11 ′ and ND12 ′ is S1, and the similarity between ND12 ′ and ND2 ′ is S2, and when S2> S1, the test mode is simulated when acquiring ND2 ′. It is possible to diagnose that preventive maintenance is necessary by the next periodic maintenance because the corresponding aging has progressed. In addition, by evaluating the relative magnitude relationship between S1 and S2, it is possible to diagnose the progress of the secular change (not progressing so much, progressing rapidly, etc.).

  According to the diagnostic method of the power conversion device according to the fifth embodiment, the neural network configuration data NDT related to the diagnostic template, the first neural network configuration data ND11 ′ at the first time t1, and the first at the second time t2. Diagnosing the operation status of the first power conversion device 110 based on the neural network configuration data ND12 ′, in other words, diagnosing a neural network at different times in the same actual machine by comparing and collating the changes in time series. Thereby, the progress degree of the secular change of the part which comprises the 1st power converter device 110 can be estimated, and it can contribute to preventive maintenance.

  In addition, since the diagnostic method of the power converter according to the fifth embodiment has the same configuration as the diagnostic method of the power converter according to the first embodiment except for the configuration described above, the diagnostic method of the power converter according to the first embodiment It has the corresponding effect as it is.

  Note that FIG. 16 does not describe how diagnosis is performed using the map and how diagnosis is performed using the reference database. However, in the fifth embodiment, the diagnosis using the map and the diagnosis using the reference database in the third and fourth embodiments can be similarly applied, and the power conversion devices according to the second to fourth embodiments can be applied. It has the same effect as the diagnostic method has.

[Embodiment 6]
Hereinafter, an example of a system that implements the diagnostic method for the power conversion device according to the present invention will be described as a sixth embodiment.

  As shown in FIG. 17, the power conversion device diagnosis system for implementing the power conversion device diagnosis method according to the present invention uses a neural network for the operation status of the first power conversion device 110 included in the actual machine 100. A power conversion device diagnosis system for performing diagnosis by using (i) a first power conversion device 110 and (ii) a first neural network that generates an entity of first neural network configuration data using a first neural network. A plurality of network configuration data generation units and physical quantities related to the first power converter are measured to obtain a first measurement value, and the first neural network configuration data is updated based on the first measurement value A first learning device 120 including a first learning unit that performs first learning repeatedly, and (iii) an external and at least a first neural A real machine 100 each having a communication device 180 for communicating network configuration data, (iv) a second power conversion device 210 having the same configuration as the first power conversion device 110, and (v) (a) a first neural A second neural network configuration data generating unit that generates a second neural network configuration data entity according to the i-th test mode (i is a natural number) using a network; and (b) a second power converter using a second power converter. While performing a simulated operation in the i test mode, a physical quantity related to the second power converter 210 is measured to obtain a second measurement value, and the second neural network according to the i test mode is obtained based on the second measurement value. A second learning unit (represented by the second learning device 220 in FIG. 17) performs the second learning by repeating updating the network configuration data a plurality of times. And (c) a reference database storage unit (not shown) that stores, in the reference database (reference DB), the second neural network configuration data related to the i-th test mode in which the second learning is reflected. , I is executed m times while increasing the number of i from 1 to m (m is an integer greater than or equal to 2), and a reference database (reference DB) comprising m second neural network configuration data is constructed A simulation operation device 200 having a database construction unit (not shown), and (vi) a test mode for performing evaluation diagnosis from m second neural network configuration data stored in the reference database Neural network configuration data corresponding to the at least one of the neural network configuration data related to the diagnostic template A reference database access unit 320 that selects one, (vii) a communication unit 310 that communicates at least the first neural network configuration data with the outside, and (viii) neural network configuration data and first learning related to the diagnostic template. The power conversion device diagnosis system includes an evaluation diagnosis device 300 having a diagnosis unit 330 for diagnosing the operation status of the first power conversion device based on the reflected first neural network configuration data.

  The evaluation diagnostic apparatus 300 may be realized by a PC, for example. Further, a reference database (reference DB) that is a part of the simulated driving apparatus may be constructed on a PC.

  Further, the exchange of information such as the first neural network configuration data between the actual machine 100 and the evaluation / diagnosis apparatus 300 may be performed by direct communication between the two, or, as shown in FIG. It may be via communication. Further, the communication may be performed using a dedicated line, or may be performed using a cloud as indicated by CMM in FIG. Further, it may be wireless or wired. Moreover, the exchange via media, such as CD-ROM, may be sufficient.

  Furthermore, the simulated driving apparatus 200, the evaluation diagnosis apparatus 300, and the reference database may be provided on the power conversion apparatus manufacturer side. By performing simulated operation and evaluation diagnosis on the power conversion device manufacturer side, detailed information necessary for maintenance and preventive maintenance can be obtained, and a flexible service can be provided to the customer.

[Example]
As shown in FIG. 18, a battery backup power supply device is used as the power conversion device 110, a predetermined commercial power supply, a battery, and a BAT and a load L are connected to the power conversion device 110. Carried out.
The power converter 110 used in the embodiment includes an outlet CNT1, an outlet CNT2, a terminal T1, a terminal T2, and a terminal T3. In addition, the first power conversion device 110 includes the bidirectional conversion unit 116 that converts power between the AC power source and the DC power source, and the first power conversion device 110 as a whole, including the control of the two-way conversion unit 116. While holding the first neural network configuration data ND1, a program for executing the actual machine learning step, etc., the microcomputer unit 112 for executing the program, an analog signal input from the outside as a measured physical quantity A / D unit 114 that converts the signal into a digital signal, an external interface unit (external I / F unit 115) that communicates with the outside, and between the commercial power source, the load L, and the bidirectional conversion unit 116 A switch S1 and a switch S2 that control connection / disconnection are provided.
The bidirectional conversion unit 116 converts the AC power source into a DC power source and supplies power for charging the battery BAT, and converts the DC power source into the AC power source and supplies the power discharged from the battery L to the load L. An inverter circuit 113b.
In the embodiment, a commercial power source (AC100V, single phase) is provided in CNT1, a predetermined load (electric product) is provided in CNT2, a predetermined battery BAT is provided in T1, and a sensor unit is provided in the vicinity of the battery BAT in T2. A temperature detector TH was connected to each.

  The first neural network NN1 has a hierarchical structure including an input layer (5 units), two intermediate layers (4 units each), and an output layer (2 units) as shown in FIG. Type neural network as a basic architecture, the input layer is constituted by units having the voltage and current of the battery BAT, the voltage and current of the load L, and the temperature of the battery BAT as input factors, and the discharge time of the battery BAT And the output layer was comprised by the unit which makes charging time an output factor. A sigmoid function was used as the transfer function.

  The physical quantities related to the first power converter 110 include the voltage and current of the battery BAT, the voltage and current of the load L, the temperature of the battery BAT, and the discharge time of the battery BAT corresponding to the input factor and output factor. And charging time was adopted. Among these, the physical quantity of the voltage and current of the battery BAT, the voltage and current of the load L, and the temperature of the battery BAT are actually measured values as “first measurement values”, and the discharge time of the battery BAT and As the charging time, the discharging time and the charging time ascertained by the microcomputer program were regarded as “first measured values”, and these “first measured values” were normalized and applied to the first neural network NN1.

  The first neural network configuration data ND1 is defined as a data structure corresponding to the basic architecture and implemented as software on the microcomputer unit 112 (not shown). Similarly, a program including an algorithm for performing the actual machine learning step S100 is implemented as software on the microcomputer unit 112 (not shown). Learning was performed using an error back-propagation method using the charging time and discharging time of the battery BAT as teacher signals.

  In the actual machine 100, while the load (electrical product) is continuously operated in a hot summer environment, the timing at which the battery is charged for a predetermined period (for at least one week) and conversely the power is discharged from the battery. The first learning is performed at the timing to obtain the first neural network configuration data ND1 ′.

On the other hand, the second power converter 210 having the same configuration as that of the first power converter 110 was used on the simulated operation apparatus side (not shown). Similar to the actual machine 100 side, the second power conversion device 210 stores a program including the second neural network configuration data ND2 based on the basic architecture of the first neural network NN1 and an algorithm for executing the reference database construction step S400. It was implemented as software on the microcomputer part installed in the. Learning was performed in the same manner as the first learning.
Further, a simulated environment providing jig (not shown) that can provide the second power converter 210 with the same operating environment, input / output, and the like as the actual actual machine 100 was used. The simulated environment providing jig is configured by a circuit that can repeatedly charge and discharge the battery BAT and the load L, and the condition of the battery BAT can be changed by using a DC power source (not shown). I made it.

A simulation operation was performed under each of the following test modes to perform second learning, and a plurality of second neural network configuration data ND2 ′ reflecting the second learning was obtained to construct a reference database.
(First test mode) Test mode in which a little stress was applied Switching between charging and discharging of the battery was performed more frequently than usual.
(Second test mode) Test mode to promote aged deterioration The environmental temperature was set to a high temperature that was full of the standard, and the voltage and current were swung to the full width of the standard for input and output.
(Third test mode) Test mode for operation under normal operating conditions

Next, comparison / collation by map will be described.
The actual learning was performed using the neural network shown in FIG. 19 (5 units in the input layer and 2 units in the output layer). However, in the evaluation diagnosis step S300, comparison / collation is simplified and explained. In order to facilitate the evaluation diagnosis, the evaluation diagnosis will be described with a map in a compact form rounded to the basic architecture shown in FIG. 20 (three units in the input layer and one unit in the output layer).

The map of the first neural network formed by the actual machine learning step S100 in the actual machine 100 is MAP1 as shown in FIG.
On the other hand, the maps of the second neural network corresponding to the first test mode, the second test mode, and the third test mode formed by the simulated driving learning step S200 in the simulated driving apparatus 200 are shown in FIG. ), MAP21, MAP22, and MAP23 as shown in FIGS. 21 (c) and 23 (d).

  In the first map MAP1 (FIG. 21A) on the real machine 100 side, the first unit from the top of the input layer is relatively large, the other units are relatively small, and the other units are approximately the same size. Can be grasped. In addition, the third and fourth units from the top of the second intermediate layer are relatively large. The second map on the simulated driving apparatus 200 side having the closest feature to these features (having the closest shape) is MAP22 (FIG. 21C). That is, among the second maps based on the three second neural network configuration data stored in the reference database, the map having the highest similarity with the shape of the first map MAP1 is MAP22. Therefore, it can be estimated that the first power conversion device 110 was in an operation situation similar to the second test mode (a test mode that promotes aging degradation).

  As can be seen from the above-described embodiments, according to the power conversion device diagnosis method of the present invention, the information related to the past “operation status” in the first power conversion device 110 of the actual machine 100 is represented by the first neural network configuration. It can be obtained as data ND1 ′, and based on this, the “operation status” of the first power conversion device 110 mounted on the actual machine 100 can be visualized and grasped, and diagnosis according to the situation of each actual machine Confirmed that can be done.

  As mentioned above, although this invention was demonstrated based on said each embodiment, this invention is not limited to each said embodiment. The present invention can be implemented in various modes without departing from the spirit thereof, and for example, the following modifications are possible.

(1) The application target described in the first to sixth embodiments, how to realize the constituent elements, the place where learning or diagnosis is performed, the number of learning times, and the like are examples, and the scope of the present invention is not impaired. It is possible to change in

(2) In the first to sixth embodiments, the description has been made with various physical quantities related to the battery and the load as the input factor or output factor of the first neural network in mind, but it is not limited to this. . A physical quantity different from these can also be applied as an input factor or an output factor. Further, a basic architecture different from the first neural network described in each embodiment can be used.

(3) In the first to sixth embodiments, the software for executing the actual machine learning step has been described with the configuration mounted on the CPU of the microcomputer mounted in the first power conversion device 110 in mind. However, it is not limited to this. If the first learning is appropriately performed in accordance with the operation of the actual machine 100 or the first power conversion device 110, the software is not installed inside the first power conversion device 110 but is installed on a computer outside the first power conversion device 110. Also good. For example, it may be implemented on another computer provided in the actual machine 100, or may be implemented on a computer provided outside the actual machine 100. Thus, the processing load of the microcomputer mounted in the power converter 110 can be reduced by executing the actual machine learning step by the computer provided outside.

(4) In the first to sixth embodiments, the timing at which the first learning step S120 is performed has been described by taking as an example the timing at which power is charged to the battery, and conversely the timing at which power is discharged from the battery. It is not limited to. It is also possible to intensively learn about the timing of turning on and off the power to a connected load (electrical equipment, electrical products, etc.). Since these timings are the timings at which stress is applied to various circuits constituting the power converter, it is expected that effective learning contributing to maintenance or preventive maintenance is performed.

(5) In the first to sixth embodiments, the configuration has been described in which the first learning is performed by configuring one neural network NN1 for one first power conversion device, but is not limited thereto. .
For example, when assuming a relatively large-scale system such as a power supply backup system for a factory or a power management system for a power plant, a relatively small-scale architecture is used for one large-scale first power converter 110. In the evaluation diagnosis step S300, the first neural network configuration data associated with the plurality of neural networks is sucked up, and the first neural network configuration data is acquired. It is good also as performing evaluation diagnosis by comprehensively grasping. With such a relatively small neural network, it can be expected that the convergence of the first learning is accelerated. Moreover, the learning (calculation) load in the first power conversion device 110 is reduced, and it is not necessary to secure a large amount of CPU resources for the computer mounted on the first power conversion device 110.
Conversely, a case where one large-scale actual device 100 is configured by a plurality of first power conversion devices 110 is also conceivable, but in this case, the power according to the first to sixth embodiments for each first power conversion device 110. It is also possible to apply the diagnosis method of the conversion device to perform learning, evaluation and diagnosis, and to comprehensively evaluate and diagnose these multiple diagnosis results.

(6) In Embodiments 1 to 6, the power conversion device 110 has been described with a household storage battery system, an emergency power supply, a power conditioner, and the like in mind, but is not limited thereto. The diagnostic method for the power conversion device according to the first to sixth embodiments can be applied to all power conversion devices for industrial use, home use, transportation use, and the like.

DESCRIPTION OF SYMBOLS 100 ... Actual machine, 110 ... 1st power converter, 112 ... Microcomputer part, 113a ... Converter circuit, 113b ... Inverter circuit, 114 ... A / D part, 115 ... External interface part (external I / F part), 116 ... Both Direction conversion unit, 120 ... first learning device, 180 ... communication device, 200 ... simulated driving device, 210 ... second power conversion device, 300 ... evaluation diagnostic device, 310 ... communication unit, 320 ... reference database access unit, 330 ... Diagnosis unit, 400 ... relay device, 910 ... secondary battery characteristic evaluation device, 930 ... secondary battery mounting device, BAT ... battery, L ... load, TH ... temperature detector

Claims (9)

A method for diagnosing a power conversion device that performs a diagnosis using a neural network on an operation status of a first power conversion device,
A first neural network configuration data generating step for generating an entity of the first neural network configuration data using the first neural network; and a physical quantity related to the first power converter is measured to obtain a first measured value. A first learning step that acquires and updates the first neural network configuration data based on the first measurement value a plurality of times to perform first learning;
A second neural network configuration data generating step for generating an entity of second neural network configuration data using the first neural network; and a simulation of a second power conversion device having the same configuration as the first power conversion device. Measuring the physical quantity related to the second power conversion device to obtain a second measurement value while performing the operation, and updating the second neural network configuration data based on the second measurement value a plurality of times A simulated driving learning step including a second learning step of repeatedly performing second learning;
The second neural network configuration data reflecting the second learning is used as the neural network configuration data related to the diagnostic template, and the neural network configuration data related to the diagnostic template and the first learning reflecting the first learning are reflected. An evaluation diagnosis step including a diagnosis step of diagnosing an operation state of the first power conversion device based on neural network configuration data. A method for diagnosing a power conversion device that performs a diagnosis using a neural network on an operation status of a first power conversion device,
A first neural network configuration data generating step for generating an entity of the first neural network configuration data using the first neural network; and a physical quantity related to the first power converter is measured to obtain a first measured value. A first learning step that acquires and updates the first neural network configuration data based on the first measurement value a plurality of times to perform first learning;
(A) a second neural network configuration data generating step for generating an entity of second neural network configuration data according to the i th test mode (i is a natural number) using the first neural network; While performing a simulated operation in the i-th test mode using a second power conversion device having the same configuration as the first power conversion device, a physical quantity related to the second power conversion device is measured to obtain a second measurement value A second learning step of performing second learning by repeatedly updating the second neural network configuration data according to the i-th test mode based on the second measurement value, and (c) the A reference database storage step of storing, in the reference database, second neural network configuration data related to the i-th test mode in which the second learning is reflected, i. From 1 to m (m is an integer of 2 or more), running m times while increasing the number of i, and the reference database construction step of constructing the reference database of m of said second neural network configuration data,
Among the m pieces of the second neural network configuration data stored in the reference database, at least any one of the neural network configuration data corresponding to the test mode for performing the evaluation diagnosis is used as the neural network configuration data related to the diagnostic template. Diagnostic template selection step for selecting one of the above, the neural network configuration data related to the diagnostic template, and the operation status of the first power conversion device based on the first neural network configuration data reflecting the first learning A diagnostic method for a power converter, comprising: an evaluation diagnostic step including: In the diagnostic method of the power converter device according to claim 2,
In the evaluation diagnosis step, in place of the diagnosis template selection step according to claim 2, the m neural network configuration data stored in the reference database is suitable for the diagnosis. Select n pieces of second neural network configuration data (n is an integer of 2 or more, and n ≦ m), and enter a test mode for performing evaluation diagnosis from the n second neural network configuration data. A diagnostic method for a power converter, comprising: a second diagnostic template selection step of selecting at least one of the corresponding neural network configuration data as neural network configuration data related to a diagnostic template. In the diagnostic method of the power converter device in any one of Claims 1-3,
As the neural network, a hierarchical neural network having an input layer, one or more intermediate layers and an output layer is used.
In the first learning step and the second learning step, the first learning and the second learning are performed by an error back propagation method. In the diagnostic method of the power converter device in any one of Claims 1-4,
In the diagnostic process,
A first map is created based on the first neural network configuration data reflecting the first learning, a second map is created based on the neural network configuration data related to the diagnostic template, and the pattern of the first map And comparing the pattern of the second map with each other and evaluating the similarity of the pattern, and diagnosing the operation status of the first power converter according to the similarity Method. In the diagnostic method of the power converter device in any one of Claims 1-5,
The first power converter includes at least an inverter circuit, and is connected to at least a battery and a load. In the diagnostic method of the power converter device according to claim 6,
In the first neural network, an input layer is configured by units having the voltage and current of the battery, the voltage and current of the load, and the temperature of the battery or the temperature of the first power converter as input factors. And the output layer is comprised by the unit which makes discharge time and charge time of the said battery an output factor, The diagnostic method of the power converter device characterized by the above-mentioned. In the diagnostic method of the power converter device in any one of Claims 1-7,
The method for diagnosing a power converter, wherein the simulated operation of the second power converter includes an operation under a test mode in which stress is applied to the second power converter. In the diagnostic method of the power converter device in any one of Claims 1-8,
In the actual machine learning step,
The first neural network configuration data at the first time at which the first learning in the first power conversion device has progressed is acquired and stored in association with time information, and the first learning has further progressed. Acquiring the first neural network configuration data at a second time later than the time and holding it in association with time information;
In the evaluation diagnosis step,
Operation of the first power converter based on the neural network configuration data related to the diagnostic template, the first neural network configuration data at the first time, and the first neural network configuration data at the second time A diagnostic method for a power conversion device, characterized by diagnosing the situation.