JP2007257366A - Diagnostic device and diagnostic method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To more accurately diagnose a state of an object, in a diagnostic device and a diagnostic method. <P>SOLUTION: This diagnostic device has: a detection means 1 detecting data comprising a plurality of parameters changing according to the state of the object; a normal neuron set storage means 4a storing a normal neuron set compressed by teacherless learning of a neural network with the data detected from the object in a normal state as a data set; an abnormal neuron set storage means 4b storing a plurality of abnormal neuron sets compressed with the data detected from the object in an abnormal state as a data set together with the kind of the abnormal state; and a diagnostic means 5 diagnosing the kind of the abnormal state in the object on the basis of an actual data set and the abnormal neuron sets stored in the abnormal neuron set storage means 4b with the data detected by the detection means 1 as the actual data set when the state of the object is not decided. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention uses an unsupervised learning technique of a neural network to diagnose the state of various objects such as various machines, life forms such as animals and plants, microorganisms, and natural phenomena such as weather and celestial motion. In particular, the present invention relates to a diagnostic apparatus and a diagnostic method suitable for use in diagnosing an abnormal state occurring in a work machine such as a hydraulic excavator.

  In conventional maintenance of machinery, post-maintenance that is repaired after a failure occurs in the machinery and uniform preventive maintenance based on the usage time of the machinery are generally performed. Subsequent maintenance takes a significant amount of time and cost for repairs.In preventive maintenance, because of uniform processing, unnecessary parts and oil are discarded, increasing the cost burden on customers. However, there has been a problem of increased costs due to labor-intensive types, but in the future it will be necessary to move away from such conventional maintenance and switch to predictive maintenance.

Predictive maintenance means early diagnosis of machine abnormalities by diagnosing soundness and predicting deterioration and remaining life from inference based on load / environment information during operation, past maintenance history database, understanding of failure physics, etc. Discover and provide a safe operating environment.
For example, Patent Document 1 discloses a technique for detecting an abnormality in a target plant using a self-organizing model, with regard to an abnormality detection apparatus for a fixed machine facility such as a batch plant or a continuous plant. In other words, normal data when the target plant is in a normal state is collected in advance, and features of normal data are extracted based on this normal data using a self-organizing map (Self-Organizing Map). A feature map representing the distance relationship between each output unit in the output layer of the self-organizing map is formed and stored as a normal state model, and this normal state model and unknown input data (input vector) Based on the above, an abnormality of the target plant is detected.

  The self-organizing map is one of self-organizing models applied for the purpose of reducing the dimension of a variable while retaining the characteristics of a complex multidimensional variable. For example, in the technique described in Patent Document 1, the normal state model is stored inside the abnormality detection device as a normal data that is multidimensional data visualized and converted into a two-dimensional map. Then, when the unknown input data is compared with the normal state model and is considered to have the same characteristics, it is determined that the input data is normal data. With such a configuration, comprehensive abnormality detection regarding multidimensional input data can be realized in real time.

  However, since this technique has only one normal state model for determining whether or not the input data is normal data, it is difficult to apply, for example, when there are many operation modes of the machine. In other words, since approximately the same number of clusters as these operation modes are formed in one two-dimensional self-organizing map, the area of each cluster becomes smaller and the overlap with adjacent clusters becomes stronger. The boundary will be unclear.

As described above, the technique described in Patent Document 1 cannot perform an accurate diagnosis on a machine having a plurality of normal state model features. In addition, it is difficult to predict which measurement data, that is, how much deterioration or abnormality has occurred in which parameter.
In order to solve the above-described problems, Patent Document 2 discloses a technique for determining each operation of a target object such as a machine that can operate in a plurality of operation operations (operation modes). A self-organizing map as a separation model is disclosed.

For example, in the case of a working machine such as a hydraulic excavator, “operation from the start of scooping of earth and sand with a bucket until the end of scooping (operation mode 1)”, “after scooping up earth and sand, turning the vehicle body, "Operation to transport the truck up to the top of the vessel (operation mode 2)", "Operation from opening the bucket to starting the transfer of earth and sand to the vessel (operation mode 3)", Assume four operation modes such as “operation until the bucket is accumulated to the position of the earth and sand and operation mode 1 is entered (operation mode 4)”. On the other hand, the engine speed detected from the work machine, the discharge pressure of the hydraulic pump, the oil temperature in the hydraulic circuit, the operating pressure that controls the forward / backward / turning of the vehicle body, the operating pressure of the bucket cylinder that controls the bucket, and the stick Among the parameters such as the operating pressure of the stick cylinder that controls the boom and the operating pressure of the boom cylinder that controls the boom, the parameters detected at the same time are collectively treated as one multidimensional parameter, and the multidimensional corresponding to each operation mode A self-organizing map of parameters is formed individually. With such a configuration, each operation mode of the work machine can be accurately recognized, or if it does not apply to any of the modes, it can be confirmed that the mode is an unknown mode or an abnormal mode.
JP 11-338848 A JP 2005-25351 A

  However, the technique described in Patent Document 2 merely determines that the abnormality is unknown or abnormal based on “not normal” and cannot diagnose what kind of content the abnormality is. For example, if the judgment result of the operation mode of the work machine does not apply to any of the operation modes 1 to 4, it is possible to grasp at least that it is “not normal”, but determine what action should be taken. As a result, the same preventive maintenance measures as conventional maintenance must be taken.

As described above, according to the conventional technology, although deterioration and abnormality can be predicted, the type of deterioration and abnormality cannot be predicted, and more accurate state determination is desired from the viewpoint of predictive maintenance. ing.
The present invention has been made in view of such a problem, and an object of the present invention is to provide a diagnostic apparatus and a diagnostic method that can more accurately diagnose the state of an object.

  In order to achieve the above object, a diagnostic apparatus according to claim 1 of the present invention is a diagnostic apparatus for diagnosing the state of an object, and comprises a plurality of parameters that vary according to the state of the object. A detection means for detecting data, and a normal neuron set for storing a normal neuron set compressed by unsupervised learning of a neural network, using a plurality of the data detected by the detection means from the object in a normal state as a data set A storage means and a plurality of sets of abnormal neuron sets compressed by unsupervised learning of a neural network as a data set including a plurality of the data detected by the detection means from the object in an abnormal state together with the types of the abnormal states An abnormal neuron set storage means for storing, and the detection when the state of the object is indeterminate Whether or not the object is in a normal state based on the real data set and the normal neuron set stored in the normal neuron set storage means as a plurality of the data detected by the stage And a diagnostic means for diagnosing the type of the abnormal state in the object based on the actual data set and the abnormal neuron set stored in the abnormal neuron set storage means. It is said.

  The diagnostic device of the present invention according to claim 2 is the diagnostic device according to claim 1, wherein the abnormality of the actual data set decreases as the diagnostic means decreases the Euclidean distance between the actual data set and the abnormal neuron set. Similarity determination means for determining that the similarity to the neuron set is high, and similar neuron set selection means for selecting the abnormal neuron set determined to have the highest similarity with the actual data set by the similarity determination means; And a state diagnosing unit for diagnosing the abnormal state in the object as a type of the abnormal state corresponding to the abnormal neuron set selected by the similar neuron set selecting unit.

  According to a third aspect of the present invention, there is provided the diagnostic device according to the first aspect, wherein the plurality of data detected by the detecting means when the state of the object is uncertain is actual data. As a set, comprising compression means for compressing into a real neuron set by unsupervised learning of a neural network, the diagnostic means, the smaller the Euclidean distance between the real neuron set compressed by the compression means and the abnormal neuron set, Similarity determination means for determining that the similarity of the real neuron set to the abnormal neuron set is high, and similar neuron set selection means for selecting the abnormal neuron set determined to be the highest by the similarity determination means The abnormal neuron set selected by the similar neuron set selection means as the abnormal state in the object Is characterized by having a condition diagnosis means for diagnosing that the the abnormal state of the corresponding type.

3. The diagnostic apparatus according to claim 2, wherein the similarity determination means calculates an average value of Euclidean distances between each data constituting the actual data set and each neuron constituting the abnormal neuron set. It is preferable to determine the similarity (claim 4).
3. The diagnostic apparatus according to claim 2, wherein the abnormal neuron set storage means calculates a weight for each neuron when compressed into the abnormal neuron set, and the similarity determination means constitutes the actual data set. It is preferable to determine the similarity based on the average value of the Euclidean distance between each data to be performed and each neuron constituting the abnormal neuron set and the weight.

3. The diagnostic apparatus according to claim 2, wherein the similarity determination unit calculates an average value of minimum Euclidean distances between each data constituting the actual data set and each neuron constituting the abnormal neuron set. Thus, it is preferable to determine the similarity (claim 6).
3. The diagnostic apparatus according to claim 2, wherein the abnormal neuron set storage means calculates a weight for each neuron when compressed into the abnormal neuron set, and the similarity determination means constitutes the actual data set. It is preferable to determine the similarity based on the average value of the minimum Euclidean distance between each data to be processed and each neuron constituting the abnormal neuron set and the weight.

Further, in the diagnosis apparatus according to claim 2, it is preferable that the similarity determination unit calculates the Euclidean distance between the centroid of the actual data set and the centroid of the abnormal neuron set to determine the similarity. Item 8).
The diagnostic apparatus according to claim 3, wherein the similarity determination means calculates an average value of Euclidean distances of all pairs of each neuron constituting the real neuron set and each neuron constituting the abnormal neuron set. Thus, it is preferable to determine the similarity (claim 9).

  4. The diagnostic apparatus according to claim 3, wherein the compression means calculates a first weight for each neuron when compressing to the real neuron set, and the abnormal neuron set storage means compresses to the abnormal neuron set. Sometimes the second weight for each neuron is calculated, and the similarity determination means has an average value of Euclidean distances of all pairs of each neuron constituting the real neuron set and each neuron constituting the abnormal neuron set, It is preferable that the similarity is determined based on the first weight and the second weight.

The diagnostic apparatus according to claim 3, wherein the similarity determination means calculates an average value of minimum Euclidean distances from each neuron constituting the real neuron set to each neuron constituting the abnormal neuron set, and It is preferable to determine the degree of similarity (claim 11).
The diagnostic device according to claim 3, wherein the compression means calculates a first weight for each neuron when compressed into the real neuron set, and the similarity determination means constitutes the real neuron set. Preferably, the similarity is determined based on an average value of minimum Euclidean distances from each neuron to each neuron constituting the abnormal neuron set and the first weight (claim 12).

  4. The diagnostic apparatus according to claim 3, wherein the compression means calculates a first weight for each neuron when compressing to the real neuron set, and the abnormal neuron set storage means compresses to the abnormal neuron set. Sometimes calculating a second weight for each neuron, and the similarity determination means includes an average value of minimum Euclidean distances from each neuron constituting the real neuron set to each neuron constituting the abnormal neuron set, Preferably, the similarity is determined based on the first weight and the second weight.

Further, in the diagnosis apparatus according to claim 3, it is preferable that the similarity determination unit determines the similarity by calculating a Euclidean distance between the center of gravity of the real neuron set and the center of gravity of the abnormal neuron set. Item 14).
The diagnostic device of the present invention according to claim 15 is the diagnostic device according to claim 2 or 3, wherein the diagnostic means sets a plurality of different calculation methods when determining the similarity in the similarity determination means. A similar neuron set selecting means for selecting the abnormal neuron set having the highest similarity for each of the plurality of different computing methods set by the similarity computing method setting means, The state diagnosing means diagnoses the abnormal state in the object as the abnormal state of the type corresponding to the abnormal neuron set having the largest number of selections by the similar neuron set selecting means.

The diagnostic device according to any one of claims 1 to 15, wherein the abnormal neuron set storage means updates or changes the abnormal neuron set using the real data set (claim 16). ).
In this case, it comprises an unknown abnormal state determination means for determining whether or not the state of the object corresponding to the actual data set is in an unknown abnormal state, and the abnormal neuron set storage means includes the unknown abnormal state determination Preferably, an unknown abnormal neuron set obtained by compressing the actual data set corresponding to the unknown abnormal state determined by the means by unsupervised learning of a neural network is newly stored together with the type of the unknown abnormal state (claim). Item 17).

Furthermore, in this case, it is preferable that the unknown abnormal state determination means determines whether or not the object is in an unknown abnormal state based on the Euclidean distance between the actual data set and the abnormal neuron set (claim) Item 18).
The diagnostic method of the present invention according to claim 19 is a diagnostic method for diagnosing a state of a target object, wherein a plurality of parameters that vary from the target object in a normal state according to the state of the target object And a normal neuron set memory for storing a normal neuron set obtained by compressing the normal data set by unsupervised learning of a neural network. An abnormal data set detecting step for detecting a plurality of sets of the data as an abnormal data set, and detecting a plurality of sets of the data as an abnormal data set from the object in an abnormal state; and a plurality of sets of the abnormal data Abnormal neuron set compressed by unsupervised learning of neural network An abnormal neuron set storage step for storing a plurality of sets together with the type of the abnormal state, and when the state of the object is uncertain, detecting data consisting of the plurality of parameters, A real data set detecting step for detecting as a real data set, a normal state determining step for determining whether or not the object is in a normal state based on the real data set and the normal neuron set, and the real data And a diagnostic step of diagnosing the type of the abnormal state in the object based on the set and the abnormal neuron set.

  The diagnostic method of the present invention according to claim 20 is the diagnostic method according to claim 19, wherein in the diagnostic step, the smaller the distance between the actual data set and the abnormal neuron set, the smaller the abnormal neuron of the actual data set. A similarity determination step for determining that the similarity to the set is high, and a similar neuron set selection step for selecting an abnormal neuron set determined to have the highest similarity with the actual data set in the similarity determination step; The abnormal state in the object comprises a state diagnosis step for diagnosing the abnormal state of the type corresponding to the abnormal neuron set selected in the similar neuron set selection step.

  The diagnostic method of the present invention according to claim 21 is the diagnostic method according to claim 19, further comprising a compression step of compressing the real data set into a real neuron set by unsupervised learning of a neural network, the diagnostic step comprising: In the similarity determination step for determining that the similarity between the real neuron set and the abnormal neuron set is high as the distance between the real neuron set and the abnormal neuron set is high, in the similarity determination step, the real neuron set A similar neuron set selection step for selecting an abnormal neuron set determined to have the highest degree of similarity to and the abnormal state in the object corresponding to the abnormal neuron set selected in the similar neuron set selection step A state diagnosis step for diagnosing the abnormal state of the type; It is characterized in that Ranaru.

21. In the diagnosis method according to claim 20, in the similarity determination step, an average value of Euclidean distances between each data constituting the actual data set and each neuron constituting the abnormal neuron set is calculated. It is preferable to determine the similarity (claim 22).
21. The diagnostic method according to claim 20, wherein in the abnormal neuron set storage step, a weight for each neuron is calculated at the time of compression into the abnormal neuron set, and the actual data set is configured in the similarity determination step. It is preferable to determine the similarity based on the average value of the Euclidean distance between each data and each neuron constituting the abnormal neuron set and the weight (claim 23).

21. In the diagnostic method according to claim 20, in the similarity determination step, an average value of minimum Euclidean distances between each data constituting the actual data set and each neuron constituting the abnormal neuron set is calculated. Thus, it is preferable to determine the similarity (claim 24).
21. The diagnostic method according to claim 20, wherein in the abnormal neuron set storage step, a weight for each neuron is calculated at the time of compression into the abnormal neuron set, and the actual data set is configured in the similarity determination step. It is preferable to determine the similarity based on the average value of the minimum Euclidean distance between each data and each neuron constituting the abnormal neuron set and the weight (claim 25).

In the diagnosis method according to claim 20, preferably, in the similarity determination step, the similarity is determined by calculating a Euclidean distance between the center of gravity of the actual data set and the center of gravity of the abnormal neuron set (claim). Item 26).
In the diagnosis method according to claim 21, in the similarity determination step, an average value of Euclidean distances of all pairs of each neuron constituting the real neuron set and each neuron constituting the abnormal neuron set is calculated. Thus, it is preferable to determine the similarity (claim 27).

  22. The diagnostic method according to claim 21, wherein in the compression step, a first weight for each neuron is calculated at the time of compression to the real neuron set, and compression to the abnormal neuron set is performed in the abnormal neuron set storage step. Sometimes calculating a second weight for each neuron, and in the similarity determination step, an average value of Euclidean distances of all pairs of each neuron constituting the real neuron set and each neuron constituting the abnormal neuron set, It is preferable to determine the similarity based on the first weight and the second weight.

In the diagnosis method according to claim 21, in the similarity determination step, an average value of minimum Euclidean distances from each neuron constituting the real neuron set to each neuron constituting the abnormal neuron set is calculated, and It is preferable to determine the degree of similarity (claim 29).
Furthermore, in the diagnostic method according to claim 21, in the compression step, a first weight for each neuron is stored at the time of compression to the real neuron set, and each of the real neuron sets constituting the real neuron set is stored in the similarity determination step. It is preferable to determine the similarity based on an average value of minimum Euclidean distances from neurons to each neuron constituting the abnormal neuron set and the first weight (claim 30).

  22. The diagnostic method according to claim 21, wherein in the compression step, a first weight for each neuron is calculated at the time of compression to the real neuron set, and compression to the abnormal neuron set is performed in the abnormal data set storage step. A second weight for each neuron is sometimes calculated, and in the similarity determination step, an average value of minimum Euclidean distances from each neuron constituting the real neuron set to each neuron constituting the abnormal neuron set, the first It is preferable to determine the similarity based on the weight and the second weight.

In the diagnosis method according to claim 21, in the similarity determination step, it is preferable to determine the similarity by calculating a Euclidean distance between the center of gravity of the real neuron set and the center of gravity of the abnormal neuron set. Item 32).
A diagnostic method according to a thirty-third aspect of the present invention is the diagnostic method according to the twentieth or twenty-first aspect, wherein the similarity determination step is repeated a plurality of times by setting a plurality of different calculation methods in the similarity determination step. In the similar neuron set selection step, an abnormal neuron set determined to have the highest degree of similarity with the real neuron set is repeatedly selected several times, and in the state diagnosis step, the abnormal state in the object is selected. It is characterized by diagnosing the abnormal state of the type corresponding to the abnormal neuron set having the largest number of selections in the similar neuron set selection step.

In the diagnostic method according to any one of claims 19 to 33, in the abnormal neuron set storage step, the abnormal neuron set is preferably updated or changed using the real data set (claim 34). ).
In this case, an unknown abnormal state determination step for determining whether or not the state of the object corresponding to the actual data set is in an unknown abnormal state is provided. In the abnormal neuron set storage step, the unknown abnormal state determination Preferably, an unknown abnormal neuron set obtained by compressing the actual data set corresponding to the unknown abnormal state determined in the step by unsupervised learning of a neural network is newly stored together with the type of the unknown abnormal state (claim). Item 35).

  Furthermore, in this case, it is preferable in the unknown abnormal state determination step to determine whether or not the object is in an unknown abnormal state based on the Euclidean distance between the actual data set and the abnormal neuron set. Item 36).

According to the diagnostic apparatus and diagnostic method of the present invention (Claims 1 and 19), it is possible to determine whether or not the target object is in a normal state. In addition to being in a state, it is possible to determine the type of abnormal state. Thereby, the state of a target object can be predicted more clearly and predictive maintenance becomes easy.
In addition, since all of a plurality of parameters that change in accordance with the state of the object are handled as one data, for example, a comprehensive determination can be made in comparison with a method in which each parameter is determined individually.

  According to the diagnostic apparatus and diagnostic method of the present invention (claims 2 and 20), the state of the object can be accurately diagnosed by determination based on the similarity derived from the distance between the actual data set and the neuron set. Can do. Further, since the similarity of the data set itself, which is a data group detected by the detection means, is calculated, diagnosis can be performed immediately after data acquisition (that is, semi-real time). Thereby, for example, early detection of an abnormal state of the object is easy.

  According to the diagnostic apparatus and diagnostic method of the present invention (claims 3 and 21), the determination is based on the similarity derived from the distance between the neuron sets (that is, the normal neuron set, abnormal neuron set, and real neuron set). The operating state of the object can be accurately diagnosed. Further, since the similarity of the real neuron set compressed by the compression means is calculated, the calculation amount (diagnosis load) of the similarity can be reduced, and the diagnosis speed can be increased. .

Moreover, according to the diagnostic apparatus and diagnostic method of the present invention (claims 15 and 33), the reliability of the determination result can be increased by the consensus determination based on a plurality of calculation methods.
Further, according to the diagnostic apparatus and diagnostic method of the present invention (claims 16 and 34), when a new abnormal state that does not exist in the current knowledge base occurs, a module composed of neurons representing the abnormal state is added. It is possible to replace the old module with a new one, and the diagnostic accuracy can be further improved.

  Further, according to the diagnostic apparatus and diagnostic method of the present invention (claims 17 and 35), by using an arbitrary actual data set corresponding to an uncertain state as an actual data set corresponding to an unknown abnormal state, A new abnormal neuron set can be created based on the actual data set. That is, a module corresponding to an actual unknown state detected by the detection means can be added to the knowledge base.

  Further, according to the diagnostic apparatus and diagnostic method of the present invention (claims 18 and 36), it is possible to accurately determine whether or not any actual data set is in an unknown abnormal state by determination based on the Euclidean distance. it can.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
FIGS. 1 to 11 illustrate a diagnostic apparatus according to an embodiment of the present invention. FIG. 1 is a schematic configuration diagram illustrating the overall configuration of the diagnostic apparatus. FIG. 2 relates to computation in the diagnostic apparatus. 3A to 3D are schematic diagrams for explaining the concept of similarity according to the determination in the diagnostic apparatus, and FIG. 4 is an average distance ADP ₀ . FIG. 5 is a schematic diagram for explaining the computation content of the average minimum distance AMD ₀ , FIG. 6 is a schematic diagram for explaining the computation content of the average distance ADP, and FIG. FIG. 8 is a schematic diagram for explaining the calculation contents of the minimum distance AMD, FIG. 8 is a schematic diagram for explaining the calculation contents of the weighted average minimum distance WAMD ₂ , and FIG. 9 shows the module creation control contents of the neuron set in this diagnostic apparatus. FIG. 10 is a flowchart of the first diagnostic apparatus of the present diagnostic apparatus. FIG. 11 is a flowchart showing the diagnostic control contents in the second diagnostic apparatus of the present diagnostic apparatus.

[Constitution]
[1] Overall Configuration As shown in FIG. 1, this diagnostic apparatus is intended for diagnosis of the state of a hydraulic excavator. From the first diagnostic apparatus 10 for performing diagnosis on the hydraulic excavator and the hydraulic excavator, And a second diagnostic device 20 for performing diagnosis at a service center at a remote location.

  The first diagnostic device 10 is a device mounted inside a hydraulic excavator, and includes a data detection device (detection means) 1, an ECU (electronic control device) 2, and a display device 3. The ECU 2 is a computer having a storage device (ROM, RAM, etc.) and a central processing unit (CPU) incorporating a processing program, and has a storage unit 4, a diagnosis unit (diagnostic means) 5, a compression unit as functional parts inside. Unit (compression unit) 6 and communication unit 7.

  On the other hand, the second diagnostic device 20 is a device provided outside the hydraulic excavator, and includes an ECU 12 and a display device 13. The ECU 12 is a computer similar to the ECU 2 of the first diagnostic apparatus 10, and includes a storage unit 14, a diagnostic unit (diagnostic unit) 15, and a communication unit 17 as functional parts. The second diagnostic device 20 is provided in a service center of a business office that owns or manages the hydraulic excavator in which the first diagnostic device 10 is provided.

The data detection apparatus 1 is provided with various sensors corresponding to parameters (variable elements) related to the hydraulic excavator. These sensors detect each parameter value that varies according to the operation of the hydraulic excavator for each operation mode.
The parameter detected here includes not only data directly detected from the sensor but also data obtained by processing certain detected data by calculation or the like, and obtaining the value of the corresponding parameter as an estimated value. For example, engine speed, fuel consumption, hydraulic pump pressure (one or more hydraulic pump pressures), oil temperature in hydraulic circuit, operating pressure to control forward / backward or turning of vehicle body, bucket cylinder to control bucket Operating pressure of the stick cylinder that controls the stick, operating pressure of the boom cylinder that controls the boom, and the like.

  In this diagnostic apparatus, each of these simultaneously detected parameters (meaning each parameter detected under the same condition, here, each parameter detected at the same time) is one data. (Data point) x is input to the ECU 2.

[2] Storage unit 4
Next, the internal configuration of the ECU 2 will be described. First, the storage unit 4 stores information serving as a reference for diagnosis, and includes a normal neuron set storage unit (normal neuron set storage unit) 4a and an abnormal neuron set storage unit (abnormal neuron set storage unit) 4b. Is done.

  The normal neuron set storage unit 4a is a knowledge base (Module Knowledge-Base) that stores the characteristics of a data group (a plurality of data and data sets) detected when the excavator is in a normal state. Here, a data set composed of a plurality of data in a normal state is first created, and a neuron set obtained by compressing the data set by unsupervised learning of a neural network is stored as a normal neuron set. Note that the knowledge base means a systematic collection of knowledge necessary for solving a specific problem with respect to a database in which information is simply accumulated.

  On the other hand, the abnormal neuron set storage unit 4b is a knowledge base that stores the characteristics of data groups detected when the excavator is in an abnormal state (an abnormal state in which some component deterioration or device failure is about to occur). . Here, a plurality of types of abnormal states such as engine failure, fuel pump deterioration, hydraulic pump failure, etc. are assumed as abnormal states, and a data set composed of a plurality of data detected under each abnormal state Are created, and a neuron set obtained by compressing each data set by unsupervised learning of the neural network is stored as an abnormal neuron set together with the type of abnormal state.

[3] Neuron sets stored in the storage unit 4 Here, the neuron sets stored in the normal neuron set storage unit 4a and the abnormal neuron set storage unit 4b will be described in detail.
In this embodiment, a plurality of neuron sets (normal neuron set) indicating the characteristics of the normal state of the hydraulic excavator and neuron sets (abnormal neuron set) indicating the characteristics of the abnormal state are prepared, and are obtained during actual operation of the hydraulic excavator. The state of the hydraulic excavator is diagnosed by determining to which neuron set the actual data set to be approximated. Each neuron set stored in the storage unit 4 is a reference knowledge for this determination, and is set in advance before the actual operation of the hydraulic excavator.

  First, a total of L sets of normal operation data sets and abnormal operation data sets are prepared. For example, when preparing one set of normal operation data sets and four sets of abnormal operation data sets from the first abnormal state to the fourth abnormal state, L = 5. In addition, the serial number of each operation data set is represented as f, and the state of the corresponding hydraulic excavator is represented for convenience. For example, f = 1 indicates a normal state, and f = 2 to 5 indicate a first abnormal state to a fourth abnormal state.

Next, in each operation data set, M pieces of data detected by the data detection device 1 are acquired in advance. This number is, for example, about several hundred to several thousand, and it is considered that these data represent typical dynamic characteristics of a predetermined state of the excavator. Hereinafter, for general purpose notation, the number of data acquired in each operation data set is expressed as M _f (f = 1, 2,..., L). For example, the dynamic characteristics in the normal state are represented by M ₁ data, and the dynamic characteristics in the first abnormal state to the fourth abnormal state are represented by M _{2 to} M ₅ data. It shall be represented.

Each data is a set of K parameters such as engine speed, fuel consumption, hydraulic pump pressure, etc., which are detected simultaneously. Therefore, the data x ^f _s constituting each operation data set can be expressed in the form of a K-dimensional vector as follows.

That is, it is considered that one operation data set is characterized by its state f by K types × M _f values.
In addition, L sets of such operation sets are prepared. For example, as the data x ¹ _s of operation data sets in a normal state, the hydraulic excavator is provided to detect data in a normal state, also, the data x ² _s ~x corresponding to the first abnormal state to the fourth abnormal state ^{In 5} _s , detection data obtained when the hydraulic excavator is put into a simulated failure state, test data obtained by intentionally modifying the actually detected data, and the like are prepared.

Then, for each operating data set, M _f data x ^f _s is compressed into N (N <M) neurons. Specifically, N neurons are randomly arranged in a K-dimensional space, and the features of each operational data set are learned in neurons by unsupervised learning of the neural network using M _f data x ^f _s. Let Hereinafter, for general-purpose notation, the number of neurons compressed in each operation data set is expressed as N _f (f = 1, 2,..., L).

The neuron set composed of N _f neurons after learning is a compressed data model corresponding to the state f of the hydraulic excavator, and each neuron set is denoted as Q _f . For example, neurons set corresponding to a state f = 1 (i.e., normal neurons set) to represent and Q _1, neuron set corresponding to a state f = 2 to 5 (i.e., abnormal neuron sets), respectively, Q ₂ to Q ₅ It expresses. Hereinafter, unless otherwise specified, any one of these normal and abnormal neuron sets Q _{1 to} Q ₅ is simply represented by the symbol Q. Each neuron w ^f _q constituting this neuron set Q can be expressed in the form of a K-dimensional vector as follows.

Note that, as shown in Equation 2 above, the vector of each neuron w ^f _q can be regarded as representing coordinates in the K-dimensional space.
Furthermore, each neuron w ^f _q is in the process of learning, or has been the winner for a number of data x ^f _s of the M _f-number data x ^f _s [i.e., with respect to M _f-number data x ^f _s Then, a weight (second weight) g ^f _q indicating the ratio of becoming a winning neuron is calculated based on the following Equation 3.

Each neuron set Q has the most representative information indicating the characteristics of a normal state or a certain abnormal state of the hydraulic excavator, and is stored as an individual module (compressed data model) in the storage unit 4. Each neuron w ¹ _q and weight g ¹ _q of the neuron set Q corresponding to f = 1 is stored in the normal neuron set storage unit 4a, while the neuron set Q corresponding to f = 2 to 5 is stored. The neurons w ² _{q to} w ⁵ _q and the weights g ² _{q to} g ⁵ _q are stored in the abnormal neuron set storage unit 4b.

The neuron distribution in the K-dimensional space in each neuron set Q differs in shape, size, and density, and each neuron set Q has the most appropriate information on the dynamic behavior for each state f. Each neuron set Q is also simply referred to as a “Module” because of the characteristics of the neuron set Q as a condition determination criterion. That is, the storage unit 4 as a knowledge base includes a total of L modules according to the assumed operating state f of the excavator. The optimum value of the number N _f of the neurons, the storage capacity of the storage unit 4 becomes compact, and the number of degree not interfering with the state determination, and shall be set empirically through experimentation and testing.

  Such learning of the neuron set Q is performed in advance before performing the actual work on the hydraulic excavator or separately from the actual work (in this embodiment, this is also referred to as “preliminary operation” of the hydraulic excavator). It is preferable to keep it. Therefore, for example, before the hydraulic excavator is shipped as a product, the hydraulic excavator is experimentally operated in a normal operation state (a state in which there is no abnormality) in accordance with a series of operations that will be performed after the shipment. Learning may be performed using the obtained operation data set. Alternatively, the hydraulic excavator before shipment may be intentionally operated in an abnormal operation state, and learning may be performed using the operation data set obtained as a result.

[4] Diagnostic unit 5
The diagnosis unit 5 calculates the degree of similarity by comparing the knowledge stored in the storage unit 4 in advance as described above with the data obtained from the data detection device 1 during actual operation of the hydraulic excavator, and performs a specific diagnosis of the hydraulic excavator. The similarity determination unit (similarity determination unit, similarity calculation method setting unit) 5a, similar neuron set selection unit (similar neuron set selection unit) 5b, and state diagnosis unit (state diagnosis unit) 5c are provided. It is prepared for.

First, the similarity diagnosis unit 5a distinguishes from data input from the data detection device 1 (data referred to in the formation process of the neuron set Q) during actual operation when the state of the hydraulic excavator is uncertain. hereinafter, a set of real data that) x _r, calculates the distance to each neuron set Q stored in the storage unit 4, is to determine the similarity of the (degree of closeness). For example, a set of m actual data x _r acquired in a short operation time of about 5 minutes is defined as an actual data set R (R = {1, 2,..., M}).

The distance here means the Euclidean distance between two sets in the K-dimensional space, as schematically shown in FIG. Since one neuron set Q is a set of N _f K-dimensional neurons, the actual data set R and the neuron set Q can be arranged in the same K-dimensional space. Therefore, as shown in FIG. 2, it is determined that the shorter the distance between the actual data set R and the neuron set Q, the more similar the characteristics of the actual data set R and the characteristics of the neuron set Q are. It is determined that the characteristics of the actual data set R are different from the characteristics of the neuron set Q.

  In addition, since the distribution of neurons in each neuron set Q stored in the storage unit 4 is different in shape, size, and density, if each neuron set Q is virtually arranged in the same K-order space, As shown in FIGS. 3A to 3D, different distributions are formed. Therefore, it is determined that the actual data set R is most similar to the closest neuron set Q.

Note that the Euclidean distance between each actual data (also referred to simply as data points) x _r (r = 1, 2,..., M) constituting the actual data set R and each neuron w ^f _q constituting the neuron set Q. D (r, q) can be obtained by the following equation.

なお、類似度診断部５ａは、上記の式４を用い、より具体的には５種類の手法を用いて、Ｋ次元の実データｘ_rとＫ次元の各ニューロンセットＱとの距離を演算するようになっている。以下にその５種類の手法を説明する。

The similarity diagnosing unit 5a calculates the distance between the K-dimensional actual data _xr and each K-dimensional neuron set Q using Equation 4 above, more specifically, five types of methods. It is like that. The five types of methods will be described below.

(A) Calculation of average distance ADP ₀ Average value of the distances of all pairs (combinations) between each data point x _r constituting the actual data set R and each neuron w ^f _q constituting an arbitrary neuron set Q Calculate ADP ₀ (R, Q) (mean of All Pairs Distances). This average distance ADP ₀ (R, Q) is given by Equation 6 below.

つまりこの演算では、図４に示すように、全ての実データｘ_rと全てのニューロンｗ^f _qとの組み合わせを考慮して、ユークリッド距離D(r,q)の平均値を演算することになる。実データｘ_rとニューロンｗ^f _qとの組み合わせ総数は、ｍ×Ｎ_f通りとなる。

That is, in this calculation, as shown in FIG. 4, the average value of the Euclidean distance D (r, q) is calculated in consideration of combinations of all actual data x _r and all neurons w ^f _q. . The total number of combinations of actual data x _r and neurons w ^f _q is m × N _f .

(B) Calculation of weighted average distance WADP ₀ considering weight g ^f _q All pairs (combinations) between each data point x _r constituting real data set R and each neuron w ^f _q constituting neuron set Q relative distance), and calculates the average value of those multiplied by weight g ^f _q neurons w ^f _q (the weighted _{average) WAPD 0 (R, Q)} (weighted mean of All Pairs distances). This weighted average distance WAPD ₀ (R, Q) is given by Equation 7 below.

(C) Calculation of average minimum distance AMD _{0 For} each actual data x _r constituting the actual data set R, the average value (average minimum value) of the distance to the neuron w ^f _q having the closest Euclidean distance AMD ₀ (R , Q) (Average Minimum Distance) is calculated. This average minimum distance AMD ₀ (R, Q) is given by Equation 8 below.

Here, C _rmin indicates the minimum distance from each actual data x _r to the neuron set Q, and has a value represented by the following Expression 9.

つまりこの演算では、図５に示すように、ニューロンセットＱの各ニューロンのうち、実データセットＲに近い側のニューロンのみが参照されて、距離が演算されることになる。演算にかかる実データｘ_rとニューロンｗ^f _qとの組み合わせ総数は、ｍ通りである。

That is, in this calculation, as shown in FIG. 5, among the neurons of the neuron set Q, only the neurons closer to the actual data set R are referred to calculate the distance. The total number of combinations of the actual data x _r and the neuron w ^f _{q related} to the calculation is m.

(D) Calculation of Weighted Average Minimum Distance WAMD ₀ Considering Weight g ^f _{q The} neuron w is the distance to the neuron w ^f _{q that} is closest to the Euclidean distance for each data point x _r constituting the actual data set R. mean those multiplied by weight g ^f _q of ^f _q computing (weighted average minimum) WAMD ₀ (R, Q) and (weighted average minimum Distance). This weighted average minimum distance WAMD ₀ (R, Q) is given by Equation 10 below.

ここで、q*は、各実データｘ_rから、ニューロンセットＱのニューロンq*までのユークリッド距離D(r,q*)が最小になるニューロンの番号を示す。また、ＴはニューロンセットＱの中で最小距離の演算に参加したニューロンq*の総数（Ｔ≦Ｎ）を示す。

Here, q * indicates the number of the neuron at which the Euclidean distance D (r, q *) from each actual data x _r to the neuron q * of the neuron set Q is minimized. T represents the total number (T ≦ N) of neurons q * participating in the calculation of the minimum distance in the neuron set Q.

(E) Calculation of center-of-gravity distance COGD ₀ Calculate the center of gravity of the actual data set R and the center of gravity of the neuron set Q, and calculate the Euclidean distance COGD ₀ (R, Q) (Center-of-Gravity Distance) of both centers of gravity. Calculate. This center-of-gravity distance COGD ₀ (R, Q) is given by Equation 11 below.

In each of the above five methods, the similarity diagnosis unit 5a determines the neuron set Q having the closest distance to the actual data set R as the neuron set Q having a high similarity. The determination result here is input to the similar neuron set selection unit 5b.
On the other hand, the similar neuron set selection unit 5b selects the neuron set Q determined to have the highest similarity for each of the above methods. Accordingly, five selection results are obtained here.

  Therefore, the state diagnosing unit 5c further selects the neuron set Q having the highest number of selections from the selection results in the similar neuron set selecting unit 5b. That is, here, the final judgment is made by majority decision based on five kinds of judgments by five methods. As a result, the state of the excavator supported by the most methods is diagnosed as the type of state f corresponding to the neuron set Q, and the diagnosis result is output to the display device 3. .

When the most frequently selected neuron set Q is the normal neuron set Q ₁ (f = 1) stored in the normal data storage unit 4a, the state diagnosis unit 5c diagnoses that the hydraulic excavator is in a normal state. . On the other hand, when the most frequently selected neuron set Q is the abnormal neuron set Q _f (f ≠ 1) stored in the abnormal data storage unit 4b, it is diagnosed that the hydraulic excavator is in an abnormal state, According to the state f indicated by the neuron set, what kind of abnormal state is diagnosed. For example, when f = 2, the first abnormal state is diagnosed, and when f = 5, the fourth abnormal state is diagnosed.

[5] Compression unit 6, communication unit 7
On the other hand, the compression unit 6 temporarily stores the data x _s input from the data detection device 1 and compresses the data x _s collectively for each predetermined operation time. For example, the compression unit 6 accumulates by operating the hydraulic excavator for one day. Only thousands of data x _s are compressed into n neurons. The data compression method is the same as the method for creating each neuron set Q stored in the storage unit 4 described above. Here, if the set of real neurons obtained by compressing the data x _s (actual neuron set) is P, each real neuron w _p which constitute the actual neuron set P, as shown in Equation 14 below, K It can be expressed as a vector of dimensions.

Note that the vector of each real neuron w _p can be regarded as representing coordinates in the K-dimensional space.
The compression unit 6 in the course of learning by the actual neuron W _p, for example weights indicating which becomes the winner for a number of data x _s of the m ₀ pieces of data x _s (first weight) g _p Is calculated based on Equation 15 below.

The real neurons w _p of the real neuron set P calculated by the compression unit 6 and their weights g _p are input to the communication unit 7.
The communication unit 7 is for performing wireless transmission to the outside of the hydraulic excavator, and transmits information input from the compression unit 6 to the second diagnostic device 20 provided in the service center of the office. The communication frequency in the communication unit 7 corresponds to the compression frequency in the compression unit 6, and is transmitted once a day after the operation of the excavator is completed, for example.

[6] Second diagnostic device 20
Information transmitted from the communication unit 7 of the first diagnostic device 10 is received by the communication unit 17 of the second diagnostic device 20 and input to the diagnostic unit 15. Note that the storage unit 14 of the second diagnostic device 20 has the same configuration as the storage unit 4 of the first diagnostic device 10, and the normal neuron set storage unit 14a and the abnormal neuron set storage unit 14b include normal neurons. The same neuron set as that stored in the set storage unit 4a and the abnormal neuron set storage unit 4b is stored.

[7] Diagnosis unit 15
The diagnosis unit 15 compares the knowledge stored in the storage unit 14 in advance with the information input from the communication unit 17 to calculate the similarity, and performs a specific diagnosis of the hydraulic excavator. Similar to the information diagnosis unit 5 of the first diagnosis apparatus 10, the information diagnosis unit 15 includes a similarity determination unit (similarity determination unit) 15a, a similar neuron set selection unit (similar neuron set selection unit) 15b, and a state diagnosis unit. (State diagnosis means) 15 c is configured, but the control contents in each functional unit are slightly different.

The similarity diagnosis unit 15 a calculates the distance between each real neuron w _p of the real neuron set P input from the communication unit 17 and their weights g _p and each neuron set Q stored in the storage unit 14. Determine the degree. That is, the first diagnostic device 10 determines the similarity based on the comparison between the actual data set R and each neuron set Q, but the second diagnostic device 20 compares the neurons (real neuron set P and neuron set Q). The degree of similarity is determined based on comparison with (1). Since the real neuron w _p is a set of n K-dimensional neurons, the real neuron w _p and the neuron set Q can be arranged in the same K-dimensional space, and the Euclidean distance can be measured. Therefore, similarly to the determination in the similarity diagnosis unit 5a, it is determined that the feature of the real neuron set P and the feature of the neuron set Q are more similar as the distance between the real neuron w _p and the neuron set Q is shorter.
Further, the similarity diagnosis unit 15a calculates the distance between the real neuron w _p and each neuron set Q using six kinds of methods. The six types of methods will be described below.

(A) Calculation of average distance ADP Average value ADP of distances of all pairs (combinations) between each real neuron w _p constituting real neuron set P and each neuron w ^f _q constituting arbitrary neuron set Q Calculate (P, Q) (mean of All Pairs Distances). First, since the Euclidean distance D (p, q) between the real neuron w _p and the neuron w ^f _q is given by the following expression 17, the average distance ADP (P, Q) is expressed as the following expression 18. be able to.

つまりこの演算では、図６に示すように、実ニューロンセットＰの実ニューロンｗ_pとニューロンセットＱのニューロンｗ^f _qとの全ての組み合わせを考慮して、ユークリッド距離D(p,q)の平均値を演算することになる。これらのニューロンの組み合わせ総数は、ｎ×Ｎ_f通りとなる。

That is, in this calculation, as shown in FIG. 6, the average of the Euclidean distance D (p, q) is considered in consideration of all combinations of the real neurons w _p of the real neuron set P and the neurons w ^f _q of the neuron set Q. The value will be calculated. The total number of combinations of these neurons is n × N _f .

(B) Calculation of weighted average distance WADP considering weights g _p and g ^f _q All pairs (combinations) between each real neuron w _p and each neuron w ^f _q of neuron set Q constituting real neuron set P respect to the distance), the actual neurons W _p multiplied by weights g _p and the weight g ^f _q neurons w ^f _q of the mean value weighted average distance WAPD (P, Q) (weighted mean of All Pairs distances) Calculate as This weighted average distance WAPD (P, Q) is given by Equation 19 below.

つまり、この演算では、両端のニューロンのウェイトｇ_p，ｇ^f _qが大きいほど、演算される距離が大きくなる。

That is, in this calculation, the calculated distance increases as the weights g _p and g ^f _q of the neurons at both ends are larger.

(C) Calculation of average minimum distance AMD For each real neuron w _p constituting the real neuron set P, the average value (average minimum value) AMD (P, Q) to the neuron w ^f _q having the closest Euclidean distance ) (Average Minimum Distance) is calculated. This average minimum distance AMD (P, Q) is given by Equation 20 below.

ここでＣ_pminは、各実ニューロンｗ_pからニューロンセットＱまでの最小距離を示し、以下の式２１で示される値を持つ。

Here, C _pmin indicates the minimum distance from each real neuron w _p to the neuron set Q, and has a value represented by the following Expression 21.

つまりこの演算では、図７に示すように、ニューロンセットＱのニューロンのうち、各実ニューロンｗ_pにとって最も近いものがそれぞれ参照されて、距離が演算されることになる。演算にかかる実ニューロンｗ_pとニューロンｗ^f _qとの組み合わせ総数は、ｎ通りである。

That is, in this calculation, as shown in FIG. 7, among the neurons in the neuron set Q, the closest one to each real neuron w _p is referred to and the distance is calculated. The total number of combinations of the real neuron w _p and the neuron w ^f _q involved in the calculation is n.

(D) Calculation of Weighted Average Minimum Distance WAMD ₁ Considering Weight g _{p The} distance to the neuron w ^f _q closest to the Euclidean distance with respect to the minimum distance C _pmin from each real neuron w _p to the neuron set Q is represented by m Multiplying _p (the number of data for which the pth neuron of the real neuron set P is the winner) and multiplying by the number of operating data m ₀ , the weighted average minimum distance WAMD ₁ (P, Q) (Weighted Average Minimum Distance, version Calculate 1). This weighted average minimum distance WAMD ₁ (P, Q) is given by the following Equation 22.

前述の式１５により、ここでは各実ニューロンｗ_pからニューロンセットＱまでの最小距離Ｃ_pminに対してウェイトｇ_pが乗算されたものが算出されていることになる。

According to the above-described equation 15, here, a value obtained by multiplying the minimum distance C _pmin from each real neuron w _p to the neuron set Q by the weight g _p is calculated.

(E) Wait g _p, is multiplied by the weight g _p, g _q for the minimum distance C _pmin from g _q was considered weighted average minimum distance WAMD ₂ operations each real neuron W _p to neuronal set Q, the weighted average Calculate the minimum distance WAMD ₂ (P, Q) (Weighted Average Minimum Distance, version 2). This weighted average minimum distance WAMD ₂ (P, Q) is given by Equation 23 below.

ここで、q*は、各実ニューロンｗ_pから、ニューロンセットＱのニューロンq*までのユークリッド距離D(p,q*)が最小になるニューロンの番号を示す。つまりここでは、図８に示すように、各実ニューロンｗ_pからニューロンセットＱまでの最小距離Ｃ_pminに関して、その両端のニューロンのウェイトｇ_p，ｇ^f _qが大きいほど、演算される距離が大きくなる。

Here, q * from the actual neuron w _p, indicates the number of neurons Euclidean distance D (p, q *) to the neurons q * neuron set Q is minimized. That is, here, as shown in FIG. 8, with respect to the minimum distance C _pmin from each real neuron w _p to the neuron set Q, the greater the weights g _p and g ^f _q of the neurons at both ends, the greater the calculated distance. Become.

(F) Calculation of center-of-gravity distance COGD Calculate the center of gravity of real neuron set P and the center of gravity of neuron set Q, and calculate the Euclidean distance COGD (P, Q) (Center-of-Gravity Distance) of both centers of gravity. . This center-of-gravity distance COGD (P, Q) is given by Equation 24 below.

As described above, the similarity diagnosis unit 15a determines the neuron set Q that is closest to the real neuron set P of each real neuron w _{p as} the neuron set Q having a high similarity in each of the six types of methods. The determination result here is input to the similar neuron set selection unit 15b.
On the other hand, the similar neuron set selection unit 15b selects the neuron set Q determined to have the highest similarity for each of the above methods. Therefore, six selection results are obtained here.

  Then, the state diagnosis unit 15c selects the most frequently selected neuron set Q from the selection results in the similar neuron set selection unit 15b, and the state of the hydraulic excavator corresponds to the type of neuron set Q. A diagnosis result is output to the display device 13 after being diagnosed as f.

[flowchart]
The diagnostic apparatus according to the present embodiment is configured as described above, and the processing is executed according to the flowcharts shown in FIGS. This flowchart shows the contents of control executed in the ECUs 2 and 12, and the flow of FIG. 9 is executed in advance to set each module of the neuron set Q that becomes a diagnostic reference during the preliminary operation of the hydraulic excavator. The flow of FIGS. 10 and 11 is repeatedly executed at a predetermined cycle in order to perform diagnosis during actual operation of the hydraulic excavator.

(A) Module creation flow of neuron set Q In the flowchart shown in FIG. 9, five types are used from the normal operation data set corresponding to f = 1 to the fourth abnormal operation data set corresponding to f = 5. Neuron sets (modules) Q _{1 to} Q ₅ are created.
First, in step A10, the state type f is set to f = 1. In the subsequent step A20, M ₁ data x ¹ _s shown in the above equation 1 corresponding to the state f = 1 is detected and read. The data x ¹ _s read here is in the form of a K-dimensional vector.

On the other hand, in step A30, N ₁ neurons shown in Equation 2 above are randomly arranged in a K-dimensional space, and in subsequent step A40, N ₁ data x ¹ _s is obtained by unsupervised learning of the neural network. Compressed into neuron w ¹ _q . Thus, each neuron w ¹ _q will have appropriate information about the characteristics of the data x ¹ _s , ie the dynamic behavior for state f = 1.

In the subsequent step A50, the weight g ¹ _q of each neuron w ¹ _q obtained in the previous step is calculated based on the above equations 3 and 4. In step A60, each neuron w ¹ _q and weight g ¹ _q corresponding to the state of f = 1 is stored in the normal neuron set storage unit 4a as a neuron set (module) Q ₁ corresponding to normal operation.
Thereafter, in step A70, it is determined whether or not the state f is f = L. If f ≠ L, the process proceeds to step A80 where f + 1 is substituted for f, and the flow after step A20 is repeated. That is, when the creation of the module in the state of f = 1 is completed, the modules in the state of f = 2, the module in the state of f = 3, and the respective modules are created. Thus, each module in order normal neuronal set to Q ₁ and abnormal neuron set Q ₂ to Q ₅ states f is created and stored to the normal neuronal set storage 4a and abnormal neuronal set storage 4b.

  On the other hand, if f = L in step A70, it is considered that modules corresponding to all the operation data sets have been created, and this flow ends.

(B) Diagnosis Flow in First Diagnosis Device 10 In step B10 of the flowchart shown in FIG. 10, in the diagnosis unit 5 of the first diagnosis device 10, the data point x _r of the actual data set R input from the data detection device 1 is obtained. m are read.

In subsequent Step B20 to Step B60, the neuron set Q having a high similarity to the actual data set R is selected in the similarity determination unit 5a in each step, and five types of selection results are accumulated and stored.
First, in step B20, the average value ADP ₀ (R, Q) of the distances of all pairs between each data point x _r and each neuron w ^f _q constituting each neuron set Q is calculated based on the above equation 6. Calculated. That is, here, the average value ADP ₀ (R, Q) of the actual data set R and the neuron set Q is calculated for the _five neuron sets Q _{1 to} Q ₅ corresponding to the state f of the hydraulic excavator. Become. Then, the similarity determination unit 5a selects the neuron set Q determined to have the highest similarity in this calculation.

Subsequently, in step B30, based on the above equation 7, for all pairs of distances between each data point x _r constituting the actual data set R and each neuron w ^f _q constituting the neuron set Q, neurons W ^f _q of weights g ^f mean WAPD ₀ but _q multiplied by the (R, Q) is calculated. That is, here, the similarity between the actual data set R and the neuron set Q is measured by a method different from the previous step. In addition to the selection result in the previous step by the similarity determination unit 5a, the neuron set Q determined to have the highest similarity is selected.

In step B40, the average value AMD ₀ (R, Q) of the distance to the neuron w ^f _q having the closest Euclidean distance is obtained for each data point x _r constituting the actual data set R based on the above equation 8. Calculated. As in the previous step, the neuron set Q determined to have the highest similarity is selected independently of the selection result.
In step B50, based on the equation 10 described above, the distance of the most Euclidean distance to close the neuron w ^f _q for each data point x _r constituting real data sets R, weight g of the neuron w ^f _q ^The average value WAMD ₀ (R, Q) multiplied by ^f _q is calculated and the neuron set Q is selected. In step B60, the centroid of the actual data set R and the centroid of the neuron set Q are A distance COGD ₀ (R, Q) is calculated to select the neuron set Q.

In step B70, the similar neuron set selection unit 5b selects the neuron set Q that is selected most frequently from the selection results in steps B20 to B60. Thus, in step B80, the status diagnostic section 5c, considered to previous final sorted neurons set Q to the corresponding state f in step, is most similar to the characteristic of each data point x _r, diagnosis The result is output to the display device 3, and this flow ends.

(C) Diagnosis Flow in Second Diagnosis Device 20 In step C10 of the flowchart shown in FIG. 11, in the diagnosis unit 15 of the second diagnosis device 20, each real neuron set P input via the communication units 7 and 17 is displayed. Each neuron w _p and weight g _p are read. In other words, the first diagnostic device 10 diagnoses the operating state of the hydraulic excavator using the raw data obtained by the data detection device 1 (data not subjected to compression processing), whereas the second diagnostic device 20 The operating state of the hydraulic excavator is diagnosed using a neuron obtained by compressing raw data obtained by the data detection device 1.

In subsequent steps C20 to C70, the neuron set Q having a high similarity with the real neuron set P is selected in the similarity determination unit 15a in each step, and six types of selection results are accumulated and stored.
First, in step C20, based on the above equation 17, all of the real neuron w _p that constitutes the real neuron set P and each neuron w ^f _q that constitutes the arbitrary neuron set Q in the similarity determination unit 15a. The average value ADP (P, Q) of the distances of the pairs is calculated. Then, the neuron set Q determined to have the highest similarity in this calculation is selected.

Next, in step C30, the real neuron w is calculated for all pairs of distances between each real neuron w _p of the real neuron set P and each neuron w ^f _q of the neuron set Q based on the above equation 19. _p weights g _p and neuronal w ^f _q of weights g ^f _q mean those multiplied by WAPD (P, Q) is calculated, the neuron set Q is selected to have the highest degree of similarity.
Also in the subsequent steps C40 to C70, the average minimum distance AMD (P, Q), the weighted average minimum distance WAMD ₁ (P, Q), and the weighted average minimum distance WAMD ₂ (P , Q) and the center-of-gravity distance COGD (P, Q) are calculated, and the neuron set Q determined to have the highest similarity in each method is selected.

In step C80, the similar neuron set selection unit 51b selects the neuron set Q that is selected most frequently from the selection results in steps C20 to C70. Thereby, in step C90, the state diagnosis unit 15c regards the state f corresponding to the neuron set Q finally selected in the previous step as being most similar to the feature of each real neuron w _p , and diagnoses it. The result is output to the display device 13, and this flow ends.

[effect]
The diagnostic device according to an embodiment of the present invention has the following effects.
First, in the first diagnostic device 10 provided in the hydraulic excavator, the actual data set R composed of the raw actual data obtained by the data detection device 1 is compared with the neuron set Q stored in the storage unit 4. Therefore, the diagnosis can be performed quickly after the data is acquired. In other words, semi-real time diagnosis is possible. For example, if the display device 3 is installed in the cab of the excavator, the diagnosis result can be notified to the operator (operator) of the excavator. Maintenance maintenance can be promoted. Such semi-real-time diagnosis is considered effective when diagnosing abnormal conditions such as the hydraulic system, engine oil temperature, hydraulic oil temperature, etc., which have relatively quick state changes.

  On the other hand, the second diagnostic device 20 provided in the service center makes a diagnosis by comparing a real neuron set P composed of compressed real neurons with a neuron set Q stored in the storage unit 14. Therefore, the amount of communication data between the communication units 7 and 17 can be greatly reduced, and the communication cost for diagnosis can be reduced. In particular, the diagnosis by the second diagnosis device 20 is advantageous when it is desired to make a diagnosis that does not require immediateness in the diagnosis result, such as a monthly inspection or an annual inspection of a hydraulic excavator. For example, as an abnormal state, there is a case of diagnosing an abnormal state whose state change is relatively slow, such as when diagnosing the life of an in-vehicle battery or diagnosing a filter cleaning state of an air conditioner.

In any of the above diagnostic apparatuses 10 and 20, since the same neuron set Q is stored in the storage units 4 and 14, not only the configuration is simple, but also the knowledge base as a diagnostic standard is standardized. Can improve the credibility of the diagnosis.
In addition, since each neuron set Q stored in the storage units 4 and 14 is modularized for each state f set in advance, replacement, replacement, and addition are extremely easy. In this embodiment, the configuration is simplified and the case of diagnosing one normal state and four abnormal states is described. However, there is no limit to the number of states that can be assumed for both normal states and abnormal states. The base can be increased without limit.

  Further, according to the present diagnostic apparatus, it is possible not only to determine whether or not the hydraulic excavator is in a normal state, but also to diagnose the type of abnormality when it is in an abnormal state. That is, in the conventional diagnostic device, the abnormal state is grasped by at least “not normal”, but in this diagnostic device, it is specifically “first abnormal state” or “fourth abnormal state”. I can understand that. Further, as described above, by enhancing the knowledge base, the types of deterioration and abnormal states that can be diagnosed by the diagnostic apparatus can be increased without limit. Therefore, the state of the hydraulic excavator can be grasped more clearly, maintenance according to the diagnosed abnormal state is possible, and predictive maintenance is facilitated.

Also, with regard to the diagnostic method, since all of a plurality of parameters that vary depending on the state of the hydraulic excavator are handled as one data, for example, a summary that is practically difficult with a method in which each parameter is individually determined, for example. Judgment is possible.
Further, by capturing a plurality of parameters detected at the same time as “points” indicating a certain state, it is possible to re-recognize the distance between a set of points as “state proximity”. In addition to this, regarding the diagnosis algorithm, the similarity between the two sets is determined by the distance between the set (actual data set R or real neuron set P) based on the detected data and the neuron set Q. Therefore, it is possible to accurately grasp the characteristics of the set based on the detection data.

Moreover, in the state diagnosis part 5c, since the final judgment is made by majority based on five kinds of judgments by five kinds of methods, the reliability of the judgment result can be improved.
In this diagnostic apparatus, there are five types of calculation methods for determining the degree of similarity in the first diagnostic apparatus 10 and six types in the second diagnostic apparatus 20, and these calculation methods are likely to diagnose each other ( (Reliability, reliability) is not something that can be given superiority or inferiority.

[Others]
Although one embodiment of the present invention has been described above, the present invention is not limited to this embodiment, and various modifications can be made without departing from the spirit of the present invention.
For example, in the above-described embodiment, the storage unit 4 is configured to store in advance a module serving as a reference for diagnosis through the process illustrated in FIG. 9. For example, the storage unit 4 can be rewritten. If configured with memory, when a new abnormal condition that does not exist in the current knowledge base occurs, it is possible to add a module consisting of neurons that represent the abnormal condition, or to replace the old module with a new one. The diagnostic accuracy can be further improved.

  When updating or changing such a module, it is preferable to include means for determining whether or not the state corresponding to the actual data set is in an “unknown abnormal state” (unknown abnormal state determination unit). For example, based on the Euclidean distance between the actual data set and the abnormal neuron set, when the actual data set is not similar to any abnormal neuron set (the Euclidean distance between the actual data set and each abnormal neuron set is In addition, a determination unit that determines that the target object is in an unknown abnormal state may be provided in the ECUs 2 and 12 when the distance is larger than a predetermined distance set in advance, or an actual data set may be stored. The “unknown abnormal state” may be determined based on the judgment of the observing operator or administrator.

In any case, a new abnormal neuron set based on the actual data set is created by using an arbitrary actual data set corresponding to the undetermined state as an actual data set corresponding to the “unknown abnormal state”. be able to. That is, a module corresponding to an actual unknown state detected by the detection means can be added to the knowledge base.
In the above-described embodiment, the similarity determination unit 5a calculates the distance between the K-dimensional actual data set R, the actual neuron set P, and the K-dimensional neuron set Q using five types of methods. However, the calculation method of these distances is not limited to this. The same applies to the six types of techniques in the similarity determination unit 15a. Conversely, the distance between sets may be calculated using at least one of these methods.

  Further, in the above-described embodiment, the state diagnosis unit 5c further selects the neuron set Q that is selected most frequently from the selection results in the similar neuron set selection unit 5b. The method is not limited to this. For example, a configuration in which the sum of the distances calculated by each method is the smallest may be selected, or the reliability of the selection result may be weighted for each method.

  In the present embodiment, a hydraulic excavator is taken as an example of a diagnosis target, but the target is not limited to this, for example, vehicles such as trucks and buses, ships, and industrial machines. The present invention can be widely applied to the determination of the quality of operations of various machinery, and the like, and can also be applied to the determination of the quality of life forms such as animals and plants and microorganisms, and estimation of changes in celestial bodies such as weather or the earth.

It is a block block diagram which shows the whole structure of the diagnostic apparatus concerning one Embodiment of this invention. It is a schematic diagram for demonstrating the concept of the Euclidean distance which concerns on the calculation in this diagnostic apparatus. It is a schematic diagram for demonstrating the concept of the similarity which concerns on the determination in this diagnostic apparatus. It is a schematic diagram for explaining the operation contents of the average distance ADP ₀ in the diagnostic device. It is a schematic diagram for explaining the operation contents of the average minimum distance AMD ₀ in the diagnostic device. It is a schematic diagram for demonstrating the calculation content of the average distance ADP in this diagnostic apparatus. It is a schematic diagram for demonstrating the calculation content of the average minimum distance AMD in this diagnostic apparatus. Is a schematic diagram for explaining the operation contents of the weighted average minimum distance WAMD ₂ in the diagnostic device. It is a flowchart which shows the module creation control content of the neuron set in this diagnostic apparatus. It is a flowchart which shows the diagnostic control content in the 1st diagnostic apparatus of this diagnostic apparatus. It is a flowchart which shows the diagnostic control content in the 2nd diagnostic apparatus of this diagnostic apparatus.

Explanation of symbols

1 Data detection device (detection means)
2,12 ECU
3,13 Display device 4,14 Storage unit 4a, 14a Normal neuron set storage unit (normal neuron set storage means)
4b, 14b Abnormal neuron set storage unit (abnormal neuron set storage means)
5,15 Diagnosis section (diagnostic means)
5a, 15a Similarity determination unit (similarity determination means, similarity calculation method setting means)
5b, 15b Similar neuron set selection unit (similar neuron set selection means)
5b, 15c State diagnosis unit (state diagnosis means)
6 Compression unit (compression means)
7, 17 Communication unit 10 First diagnostic device 20 Second diagnostic device P: Real neuron set Q: Neuron set (module)
Q ₁ : Normal neuron set Q _{2 to} Q ₅ : Abnormal neuron set R: Actual data set

Claims (36)

A diagnostic device for diagnosing the state of an object,
Detection means for detecting data consisting of a plurality of parameters that vary according to the state of the object;
Normal neuron set storage means for storing a normal neuron set compressed by unsupervised learning of a neural network as a data set of a plurality of the data detected by the detection means from the object in a normal state;
An abnormal neuron that stores a plurality of abnormal neuron sets, which are compressed by unsupervised learning of a neural network, together with the types of abnormal states, as a data set, a plurality of the data detected by the detecting means from the object in an abnormal state Set storage means;
A plurality of the data detected by the detection means when the state of the object is uncertain is used as an actual data set, and the actual data set and the normal neuron set stored in the normal neuron set storage means And determining whether or not the object is in a normal state, and based on the actual data set and the abnormal neuron set stored in the abnormal neuron set storage means, the abnormality in the object is determined. A diagnostic apparatus comprising: a diagnostic unit that diagnoses a type of a state. The diagnostic means
Similarity determination means for determining that the smaller the Euclidean distance between the actual data set and the abnormal neuron set is, the higher the similarity between the actual data set and the abnormal neuron set is;
Similar neuron set selection means for selecting an abnormal neuron set determined by the similarity determination means to have the highest similarity with the actual data set;
The state diagnostic means for diagnosing that the abnormal state in the object is the abnormal state of the type corresponding to the abnormal neuron set selected by the similar neuron set selecting means. The diagnostic apparatus according to 1. A plurality of the data detected by the detection means when the state of the object is indeterminate, as a real data set, comprising compression means for compressing into a real neuron set by unsupervised learning of a neural network;
The diagnostic means
Similarity determination means for determining that the similarity of the real neuron set to the abnormal neuron set is higher as the Euclidean distance between the real neuron set compressed by the compression means and the abnormal neuron set is smaller;
A similar neuron set selection means for selecting the abnormal neuron set determined to be the highest by the similarity determination means;
The state diagnostic means for diagnosing that the abnormal state in the object is the abnormal state of the type corresponding to the abnormal neuron set selected by the similar neuron set selecting means. The diagnostic apparatus according to 1. The similarity determination means calculates the average value of the Euclidean distance between each data constituting the actual data set and each neuron constituting the abnormal neuron set, and determines the similarity. The diagnostic device according to claim 2. The abnormal neuron set storage means calculates a weight for each neuron when compressed into the abnormal neuron set,
The similarity determination means determines the similarity based on an average value of Euclidean distance between each data constituting the actual data set and each neuron constituting the abnormal neuron set and the weight. The diagnostic apparatus according to claim 2. The similarity determination means calculates the average value of the minimum Euclidean distance between each data constituting the actual data set and each neuron constituting the abnormal neuron set, and determines the similarity. The diagnostic device according to claim 2. The abnormal neuron set storage means calculates a weight for each neuron when compressed into the abnormal neuron set,
The similarity determination means determines the similarity based on the average value of the minimum Euclidean distance between each data constituting the actual data set and each neuron constituting the abnormal neuron set and the weight. The diagnostic apparatus according to claim 2, wherein the diagnostic apparatus is characterized. 3. The diagnostic apparatus according to claim 2, wherein the similarity determination unit determines the similarity by calculating a Euclidean distance between the center of gravity of the actual data set and the center of gravity of the abnormal neuron set. The similarity determination means calculates the average value of the Euclidean distances of all pairs of each neuron constituting the real neuron set and each neuron constituting the abnormal neuron set to determine the similarity. The diagnostic device according to claim 3. The compression means calculates a first weight for each neuron when compressing into the real neuron set;
The abnormal neuron set storage means calculates a second weight for each neuron when compressing the abnormal neuron set,
The similarity determination means is based on an average value of Euclidean distances of all pairs of each neuron constituting the real neuron set and each neuron constituting the abnormal neuron set, the first weight, and the second weight. The diagnostic apparatus according to claim 3, wherein the similarity is determined. The similarity determination means calculates the average value of the minimum Euclidean distance from each neuron constituting the real neuron set to each neuron constituting the abnormal neuron set, and determines the similarity. The diagnostic device according to claim 3. The compression means calculates a first weight for each neuron when compressing to the real neuron set,
The similarity determination means determines the similarity based on an average value of minimum Euclidean distances from each neuron constituting the real neuron set to each neuron constituting the abnormal neuron set and the first weight. The diagnostic apparatus according to claim 3, wherein the diagnostic apparatus is characterized. The compression means calculates a first weight for each neuron when compressing into the real neuron set;
The abnormal neuron set storage means calculates a second weight for each neuron when compressing the abnormal neuron set,
The similarity determination means determines the similarity based on the average value of the minimum Euclidean distance from each neuron constituting the real neuron set to each neuron constituting the abnormal neuron set, the first weight, and the second weight. The diagnostic apparatus according to claim 3, wherein: 4. The diagnostic apparatus according to claim 3, wherein the similarity determination means determines the similarity by calculating a Euclidean distance between the center of gravity of the real neuron set and the center of gravity of the abnormal neuron set. The diagnosis means has similarity calculation method setting means for setting a plurality of different calculation methods when determining the similarity in the similarity determination means,
The similar neuron set selection means selects an abnormal neuron set having the highest similarity for each of the plurality of different calculation methods set by the similarity calculation method setting means,
The state diagnosing means diagnoses the abnormal state in the object as the abnormal state of the type corresponding to the abnormal neuron set having the largest number of selections in the similar neuron set selecting means, The diagnostic device according to claim 2 or 3. The diagnostic apparatus according to any one of claims 1 to 15, wherein the abnormal neuron set storage means updates or changes the abnormal neuron set using the actual data set. An unknown abnormal state determination means for determining whether or not the state of the object corresponding to the actual data set is in an unknown abnormal state;
The abnormal neuron set storage unit compresses the unknown abnormal neuron set obtained by compressing the real data set corresponding to the unknown abnormal state determined by the unknown abnormal state determination unit by unsupervised learning of a neural network. The diagnostic apparatus according to claim 16, wherein the diagnostic apparatus newly stores the status type. The unknown abnormal state determination means determines whether or not the object is in an unknown abnormal state based on a Euclidean distance between the actual data set and the abnormal neuron set. The diagnostic device described. A diagnostic method for diagnosing the state of an object,
A normal data set detection step for detecting data consisting of a plurality of parameters that vary depending on the state of the target object from the target object in a normal state, and detecting a plurality of the data as a normal data set;
A normal neuron set storage step for storing a normal neuron set obtained by compressing the normal data set by unsupervised learning of a neural network;
An abnormal data set detection step for detecting a plurality of sets of the data as an abnormal data set while detecting data consisting of the plurality of parameters from the object in an abnormal state;
An abnormal neuron set storage step of storing a plurality of abnormal neuron sets obtained by compressing a plurality of abnormal data sets by unsupervised learning of a neural network together with the types of the abnormal states;
An actual data set detection step of detecting data consisting of the plurality of parameters when the state of the object is uncertain, and detecting the plurality of the data as an actual data set;
A normal state determination step for determining whether or not the object is in a normal state based on the actual data set and the normal neuron set;
A diagnostic method comprising: a diagnostic step of diagnosing the type of the abnormal state in the object based on the actual data set and the abnormal neuron set. The diagnostic step comprises
A similarity determination step for determining that the smaller the distance between the actual data set and the abnormal neuron set, the higher the similarity between the actual data set and the abnormal neuron set;
In the similarity determination step, a similar neuron set selection step of selecting an abnormal neuron set determined to have the highest similarity with the actual data set;
The state diagnosis step comprises diagnosing the abnormal state in the object as a type of the abnormal state corresponding to the abnormal neuron set selected in the similar neuron set selection step. 19. The diagnostic method according to 19. A compression step of compressing the real data set into a real neuron set by unsupervised learning of a neural network;
The diagnostic step comprises
A similarity determination step for determining that the smaller the distance between the real neuron set and the abnormal neuron set, the higher the similarity between the real neuron set and the abnormal neuron set;
In the similarity determination step, a similar neuron set selection step of selecting an abnormal neuron set determined to have the highest similarity with the real neuron set;
The state diagnosis step comprises diagnosing the abnormal state in the object as a type of the abnormal state corresponding to the abnormal neuron set selected in the similar neuron set selection step. 19. The diagnostic method according to 19. In the similarity determination step, the similarity is determined by calculating an average value of Euclidean distances between each data constituting the actual data set and each neuron constituting the abnormal neuron set. The diagnostic method according to claim 20. In the abnormal neuron set storage step, a weight for each neuron is calculated when compressing the abnormal neuron set,
In the similarity determination step, the similarity is determined based on an average value of Euclidean distance between each data constituting the actual data set and each neuron constituting the abnormal neuron set and the weight. The diagnostic method according to claim 20. In the similarity determination step, the similarity is determined by calculating an average value of minimum Euclidean distances between each data constituting the actual data set and each neuron constituting the abnormal neuron set, The diagnostic method according to claim 20. In the abnormal neuron set storage step, a weight for each neuron is calculated when compressing the abnormal neuron set,
Determining the similarity based on the average value of the minimum Euclidean distance between each data constituting the actual data set and each neuron constituting the abnormal neuron set and the weight in the similarity determination step; 21. The diagnostic method according to claim 20, characterized in. 21. The diagnosis method according to claim 20, wherein in the similarity determination step, the similarity is determined by calculating a Euclidean distance between the center of gravity of the actual data set and the center of gravity of the abnormal neuron set. In the similarity determination step, the similarity is determined by calculating an average value of Euclidean distances of all pairs of each neuron constituting the real neuron set and each neuron constituting the abnormal neuron set. The diagnostic method according to claim 21. In the compression step, a first weight for each neuron is calculated when compressing the real neuron set;
In the abnormal neuron set storage step, a second weight for each neuron is calculated at the time of compression to the abnormal neuron set,
In the similarity determination step, based on an average value of Euclidean distances of all pairs of each neuron constituting the real neuron set and each neuron constituting the abnormal neuron set, the first weight, and the second weight, The diagnosis method according to claim 21, wherein the similarity is determined. In the similarity determination step, the similarity is determined by calculating an average value of minimum Euclidean distances from each neuron constituting the real neuron set to each neuron constituting the abnormal neuron set, The diagnostic method according to claim 21. In the compression step, storing a first weight for each neuron when compressing to the real neuron set;
Determining the similarity based on the average value of the minimum Euclidean distance from each neuron constituting the real neuron set to each neuron constituting the abnormal neuron set and the first weight in the similarity determining step; The diagnostic method according to claim 21, wherein the diagnostic method is characterized. In the compression step, a first weight for each neuron is calculated when compressing the real neuron set;
In the abnormal data set storing step, a second weight for each neuron is calculated at the time of compression to the abnormal neuron set,
In the similarity determination step, based on the average value of the minimum Euclidean distance from each neuron constituting the real neuron set to each neuron constituting the abnormal neuron set, the first weight, and the second weight, the similarity The diagnosis method according to claim 21, wherein: is determined. The diagnosis method according to claim 21, wherein, in the similarity determination step, the similarity is determined by calculating a Euclidean distance between the center of gravity of the real neuron set and the center of gravity of the abnormal neuron set. In the similarity determination step, by setting a plurality of different calculation methods, the similarity determination is repeated a plurality of times,
In the similar neuron set selection step, an abnormal neuron set determined to have the highest degree of similarity with the real neuron set is repeatedly selected several times,
In the state diagnosis step, the abnormal state in the object is diagnosed as the abnormal state of the type corresponding to the abnormal neuron set having the largest number of selections in the similar neuron set selection step. The diagnostic method according to claim 20 or 21. The diagnostic method according to any one of claims 19 to 33, wherein in the abnormal neuron set storage step, the abnormal neuron set is updated or changed using the actual data set. An unknown abnormal state determination step for determining whether or not the state of the object corresponding to the actual data set is in an unknown abnormal state;
In the abnormal neuron set storage step, an unknown abnormal neuron set obtained by compressing the real data set corresponding to the unknown abnormal state determined in the unknown abnormal state determination step by unsupervised learning of a neural network is converted into the unknown abnormal neuron set. 35. The diagnosis method according to claim 34, wherein the diagnosis method is newly stored together with the type of state. 36. In the unknown abnormal state determination step, it is determined whether or not the object is in an unknown abnormal state based on a Euclidean distance between the actual data set and the abnormal neuron set. The diagnostic method described.

Images