JPH08278901A - Diagnostic device using decision tree type diagnostic knowledge - Google Patents

Abstract

PURPOSE: To eliminate misinput of data, decrease the workload, and obtain a diagnostic result in a short time by providing a learning data generation processing means and a decision tree generation processing means. CONSTITUTION: A decision tree reconstructing device 56 is equipped with a learning data generation processing part 56a which generates learning data for decision tree reconstruction from decision tree data, diagnostic instance data, and suspicious diagnostic instance data, a decision tree generation processing part 56b which generates an intermediate decision tree from the learning data by utilizing learning algorithm, and a learning data pruning processing part 56c which pruns the learning data generated by the processing part 56a by using the intermediate decision tree generated by the processing part 56b. Consequently, even when inspection items (attribute) increase in a diagnosis period, the diagnostic order can be optimized by applying the learning algorithm, and then the decision tree can automatically be reconstituted in a short time without any human intervention.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION The present invention relates to a communication system,
The present invention relates to a diagnostic device that performs diagnostic work on a medical system, a plant system, and the like by using diagnostic knowledge in a decision tree format.

[0002]

2. Description of the Related Art A diagnostic device (diagnostic type expert system) for assisting a diagnostic work such as a failure of a mechanical device by utilizing the knowledge of an expert has already been put into practical use. A prior art example of this type of diagnostic device in which the diagnostic knowledge is constructed in the form of a decision tree is shown schematically in FIG.

In the figure, 10 is a terminal device, and when a request for diagnosis is input from this terminal device 10, the diagnostic device carries out the following diagnostic operation. First, information necessary for diagnosis is received from the terminal device 10 in a conversational manner, the decision tree stored in the decision tree data memory 11 is referred to for diagnosis, and the conclusion is reported to the terminal device 10. Further, the diagnosis process is stored as diagnosis case data in the diagnosis case data memory 14, and if there is a display request from the terminal device 10, the diagnosis case is displayed via the history management unit 15.

The event analysis unit 12 controls the exchange of data between the terminal device 10 and the history management unit 15 and the inference control unit 13, which are required processing units of this device, and receives an inference request from the terminal device 10. Then, the data is passed to the inference control unit 13. The inference control unit 13 uses the decision tree data memory 11
The optimal diagnosis is selected by referring to the decision tree stored in. When the diagnosis is completed, the history management unit 15 uses the observation values and conclusions of the diagnosis as diagnosis case data and the diagnosis case data memory 1
4. If there is a request to display the diagnosis case, the diagnosis case is output.

The decision tree is generated by the decision tree generator 16 in the decision tree generator. The decision tree generation unit 16 receives a diagnosis case stored in the diagnosis case data memory 14 as an input, generates a decision tree from the diagnosis case, and outputs the decision tree data memory 1
Store in 1. However, in order to generate a decision tree that minimizes the average number of examinations for the entire diagnostic case, the decision tree generator 16 is given observation values for all examinations in the diagnostic series that make up the diagnostic case. It consists of learning algorithms applicable to

[0006]

In order to improve the diagnostic efficiency, this type of diagnostic device, which manages diagnostic knowledge in the form of a decision tree as an "ordered test sequence", makes a conclusion with a minimum number of tests. It is necessary to optimize the inspection order so that the number of inspections is reached (the average number of inspections is minimized). When the optimization of the inspection order is performed by using the conventional learning algorithm, the following problems occur.

That is, the diagnostic cases stored in the diagnostic case data memory 14 are a collection of diagnostic cases represented by a set of observation values and conclusions of the test, as shown in FIG. In diagnosing each case, conclusions are reached only with tests of the type known at the time. For this reason, a new type of inspection is generally added as the cases are collected, and it is difficult to collect cases by fixing the type of inspection to a fixed number from the beginning. Therefore, for example, inspections B, C, and D in case 1
As shown in (3), some of the observed values are unobserved (indicated by "-"). In this way, the diagnostic cases defined as the test series from the start of diagnosis to the conclusion always include the untested part for which the observed value is not given, and the previously proposed learning algorithm is applied to test. The order cannot be optimized.

As described above, since the decision tree cannot be automatically reconstructed by utilizing the diagnosis case obtained from the process of making the diagnosis, the decision tree is not reconstructed in the process of adding or changing the function of the diagnostic device. In order to make the configuration, it was not possible to make corrections and correction trials efficiently, except to manually select cases that consisted of appropriate inspections.

Therefore, the present invention applies a learning algorithm to a decision tree reconstructing method that can optimize a test order even for a diagnostic case having an unobserved part of the test caused by an increase in the types of tests during the collection of diagnostic cases. A diagnostic device using a configuration method is provided.

[0010]

According to the present invention, a first storage means for storing decision tree data of a test sequence defined in the form of a decision tree and a second storage means for accumulating diagnostic case data are provided. The accumulating means, the diagnosing means for diagnosing the new case by the decision tree from the first accumulating means, and the diagnosis case data from the second accumulating means when the diagnosis of the new case is wrong at the terminal node, If a new case has a new attribute,
The first pseudo-observed value indicating a value other than the corresponding observed value of the new case is given as the observed value of the corresponding examination of the existing diagnosis case classified within the extracted examination series, and the existing existing cases classified outside the applicable examination series are given. Learning data generation processing means for outputting, as learning data, a test sequence in which an unknown observation value does not exist by giving a second pseudo observation value indicating an arbitrary value within the range to the observation value of the relevant inspection of the diagnosis case. A decision tree generation processing means for reconstructing a new decision tree by applying a learning algorithm to the learning data from the learning data generation processing means,
There is provided a diagnostic device including learning data pruning processing means for reducing learning data by a generated decision tree.

In order to apply the learning algorithm to a diagnostic case having an undiagnosed portion, a pseudo observation value is obtained from a known case for a possible value of the uninspected portion of each case, and a case set is generated.

In one embodiment of the present invention, when the above-mentioned diagnostic means does not have a value corresponding to the observation value of the new case in the inspection item in the node before reaching the terminal node, and the diagnosis becomes impossible, the corresponding observation value is detected. If there is an arc corresponding to the first pseudo-observed value created from values other than, proceed with diagnosis along that arc, and if not, create an arc corresponding to the observed value of the new case and set the terminal node. It is a diagnostic means to be created.

In one embodiment of the present invention, in the case where the above-mentioned learning data generation processing means is an existing inspection item, the value of the corresponding inspection item of the existing diagnosis case classified in the extracted diagnosis series is used as a second pseudo value. If there is an observed value, give the first pseudo-observed value that indicates a value other than the corresponding observed value of the new case, and if it has the first simulated observed value, indicate that the corresponding observed value of the new case is also excluded. It is a learning data generation processing means for giving a first pseudo observation value.

In one embodiment of the present invention, the above-mentioned learning data pruning processing means has a pseudo observation value in the inspection result in a diagnosis case in which all the inspections are branched by the observation value up to the terminal node. If there is a diagnostic case that has an observation value in the test result if there is another in the test series, the diagnostic case that has the pseudo observation value is deleted, and the one that branches to the pseudo observation value in the inspection up to the terminal node. In some cases, if the ratio of the number of cases after the pseudo observation value division to the number of cases before branching is less than or equal to a threshold value, learning data pruning processing means that deletes all the number of cases after the pseudo observation value division is there.

As a measure for avoiding excessive mixing of pseudo-observed values in diagnostic case data, pseudo-observed values do not appear in the diagnostic series and are actually collected for all examinations of the case having pseudo-observed values. When there is at least one case having the observed value among the diagnosis cases belonging to the diagnosis series, the diagnosis case is deleted from the diagnosis case data.

Before describing the embodiments, the principle of the present invention will be described. The terms used here are generally used when explaining the learning algorithm, and the correspondence with the terms related to the explanation of the diagnostic device and the decision tree is as follows.

The "attribute" corresponds to each check for reaching the non-terminal node in the decision tree and reaching the diagnosis result in the work of diagnosis.

The "attribute name" is a unique name given to each attribute and corresponds to the node name in the decision tree.

The “attribute value” indicates the value of each attribute or the observation result obtained in each inspection. In the decision tree, it corresponds to the arc extending from the node.

The "terminal node" corresponds to the diagnosis result when the diagnosis is performed based on the decision tree.

"Class" corresponds to the type of diagnostic result.

The "first pseudo attribute value" not (VS) is
It is a pseudo value indicating that the value is not included in the set of values VS.

The "second attribute value" * is a pseudo value indicating that all values in the value range of the attribute are possible.

The expression of the diagnostic case is as follows. First, let A and C be a set of m attributes {A
1, A2 ,. ． ． , Am} and a set of p classification classes {C1, C2 ,. ． ． , Cp}, and a set of possible values of a certain attribute Ai is called Range (Ai).

The set of cases at a certain node is called S, and the set of cases at the root node is called Sφ. At this time,

[Equation 1] Then, each diagnosis case in S is represented by m + 1-tuple <V1, V2 ,. ． ． , Vm, Ck>.

Further, in the present invention, the sum of the weights associated with the diagnostic cases included in the terminal node corresponds to the frequency of occurrence of the test sequence.

A method for deriving a decision tree from a set of diagnostic cases including unknown attributes when an attribute set expands in the process of reconstructing a decision tree will be described. What is presented here is an example set S of terminal nodes.
This is a method in which a pseudo attribute value (pseudo observation value) is assigned to an attribute whose value is undecided in L to create a new case set, and a decision tree is recreated by the following procedure.

Generation of Case Set (1) Each new case En = <V1, V2 ,. ． ． , Vm, C
For k>, if it is classified by the existing decision tree,
The terminal node is L and its case set is SL. (I) If the class Ck of En is equal to the class of L, then En
To S0. (Ii) If not, En is added to S0 and the following processing is performed. (A) Attribute A of unused En from root node to L
When k (attribute value Vi) is a new attribute, the first pseudo attribute value not indicating other than Vi for the existing case in SL.
({Vi}) is given, and the second pseudo attribute value * is given for existing cases in S0 other than SL. (B) Unused En attribute A from root node to L
When k is a known attribute, if there is an existing case in SL that has Vi as an attribute value, the attribute Ak of the existing case in SL
Do not operate on the value of. (C) If there is an existing case in SL having the first pseudo attribute value not (VS) in (b), the first pseudo attribute value indicating that it is other than Vi in addition to Vi n
Rewrite to ot (VS). In (d) and (b), if there is an existing case in SL having the second pseudo attribute value *, it is rewritten to the first pseudo attribute value not ({Vi}) that also indicates other than Vi. (2) Each case En = <V1, V2 ,. ． ． , Vm, C
When k> is classified by the existing decision tree, the attribute value Vi of En in the check attribute Ak of a certain non-terminal node
If there is no corresponding arc. (I) If there is an arc corresponding to the first attribute value made from other than Vi, select the arc and proceed with the diagnosis. (Ii) If not, create an end node having only the arc corresponding to Vi and En connected to it in the case set, and add En to S0.

Using the decision tree generation S0, the number of diagnostic cases for a specific <attribute name, attribute value, class name> combination is obtained by the following calculation method, and the next decision is made by the learning algorithm of the ID3 system. Make a tree T0 '. (1) For each attribute Ai, a value Vi, j other than the pseudo attribute
And the weight Wl, n of the case En that belongs to the class Ck
Then, the sum n (k, i, j) of N is calculated and substituted into N (k, i, j). (2) The case where the attribute Ai has the first pseudo attribute not (VS) is checked, and if there is a case where the contribution calculation to the number-of-cases memory is not calculated, the following processing is performed. (I) A set of values included in VS and Range (Ai)
If they match, Wl, n is added to N (k, i, d). Here, let d be an index corresponding to not (VS). (Ii) Otherwise, included in Range (Ai),
N for all attribute values Vi, j not included in VS
(K, i, j) is calculated by the following equation (A).

[0030]

[Equation 2]

A set of cases in a test series consisting of observation values
Pruning of FIG. 3 shows that when there is a diagnostic case in which all the tests up to the terminal node are diagnosed by a test sequence that is branched by the observed value, and the test result has a pseudo observed value, the test result has the observed value. This is an example of deleting a diagnostic case having a pseudo observation value when another diagnostic case exists in the test sequence. In FIG. 3, n0 is a first pseudo observation value (pseudo attribute value) indicating that the inspection item A2 is other than 0, and * is a second pseudo observation value indicating that the inspection item A3 may be any value within the range. It is an observed value (pseudo attribute value). The diagnosis case E1 has a pseudo observation value in the examination A2, but since E2 of the same examination series also has the observation value 1 in the examination, E1 can be deleted. Also, diagnosis case E3
Has a pseudo observation value in the test A3 that is not included in the test series, but the diagnosis case E4 also has the observation value 1 in the test, so E3 can be deleted.

A collection of cases in a test series with pseudo observations
Pruning Figure 4 case is a branch A2 with to the end node, the diagnostic case number of previous corresponding arc pseudo observations n1 (not 1) is 2 (E1 and E2), the diagnostic case number before branching Since it is 6, the ratio of the former to the latter is 2/6. At this time, if the threshold value is 1/3 or more, for example, 0.4, E1 and E2 are deleted from the diagnostic case set.

[0033]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described in detail below with reference to the accompanying drawings.

FIG. 5 is a block diagram schematically showing the configuration of an embodiment of the diagnostic apparatus according to the present invention.

In the figure, 50 is a terminal device, and 5
Reference numeral 1 is a decision tree data memory connected to this terminal device 50 via an event analysis unit 52 and an inference control unit 53, and 54 is connected to this terminal device 50 via an event analysis unit 52 and a history management unit 55. It is a diagnostic case data memory. A decision tree reconstructing device 56 is connected to the decision tree data memory 51 and the diagnostic case data memory 54, and a pseudo diagnostic case data memory 57 is connected to the decision tree reconstructing device 56.

The decision tree data memory 51 is a memory for storing the data of the test sequence defined in the decision tree format.
The diagnosis case data memory 54 is a memory for storing diagnosis cases accumulated as a result of the operation of the diagnosis apparatus, and the pseudo diagnosis case data memory 57 is a memory for storing all the pseudo diagnosis cases that have produced the current decision tree.

The decision tree reconstructing device 56 uses the decision tree data,
A learning data generation processing unit 56a that generates learning data for reconstruction of a decision tree from the diagnosis case data and the pseudo diagnosis case data, and a decision tree generation processing unit that generates an intermediate decision tree from the learning data by using a learning algorithm. 56b and a learning data pruning processing unit 56c for pruning the learning data generated by the learning data generation processing unit 56a using the intermediate decision tree generated by the decision tree generation processing unit 56b.

When a diagnosis request is input from the terminal device 50, this diagnosis device carries out the following diagnosis operation. First,
Information necessary for diagnosis is received from the terminal device 50 in a conversational manner, the diagnosis is performed by referring to the decision tree stored in the decision tree data memory 51, and the conclusion is determined.
Report to Further, the diagnosis process is stored in the diagnosis case data memory 54 as diagnosis case data, and the terminal device 5
If there is a display request from 0, the history management unit 55 shows the diagnosis case.
View through.

The event analysis unit 52 controls the exchange of data between the terminal device 50 and the history management unit 55 and the inference control unit 53 which are required processing units of this device, and receives an inference request from the terminal device 50. Then, the data is passed to the inference control section 53. The inference control unit 53 uses the decision tree data memory 51.
The optimal diagnosis is selected by referring to the decision tree stored in. When the diagnosis is completed, the history management unit 55 uses the observation values and conclusions of the diagnosis as diagnosis case data and the diagnosis case data memory 5
4. If there is a request to display the diagnosis case, the diagnosis case is output.

The decision tree can be reconstructed by the decision tree reconstructing device 56. FIG. 6 is a flowchart for explaining the operation of the learning data generation processing unit 56a in the decision tree reconstructing device 56.

First, in step S61, a new diagnosis case stored in the diagnosis case data memory 54 is diagnosed, and the flow advances to step S62. If the terminal node is reached in step S62, step S67.
Then, the process proceeds to step S66 after performing the pseudo case set generation process when reaching the terminal node shown in FIG. Step S
If the terminal node is not reached at 62, the process proceeds to step S63.

In step S63, the arc of the node in which the inspection cannot be continued is examined, and the attribute value Vx of the new diagnosis case is examined.
After checking whether there is a value corresponding to the first pseudo attribute value created from other than the above, the process proceeds to S64. If there is such an arc in step S64, step S65
Go to and continue the diagnosis along the arc, step S
Return to 62.

In step S64, if there is no one corresponding to the first pseudo attribute value created from other than the attribute value Vx of the new diagnosis case, the process proceeds to step S66.

In step S66, the new diagnosis case and its weight 1 are added to the pseudo diagnosis case data memory, and then the process proceeds to END.

FIG. 7 is a flow chart for explaining the operation of the pseudo case generation process when the end node is reached.

First, in step S71, the new diagnosis case stored in the diagnosis case data memory 54 is diagnosed. If the new diagnosis case is correctly diagnosed, the process proceeds to END. If the diagnosis result is not correct, the process proceeds to step S72.

In the next step S72, it is checked whether or not there is an unused attribute for diagnosis in the inspection item (attribute) of the new diagnosis case. In the next step S73, steps S74 to S76d are performed for all attributes not used for diagnosis.
Check whether or not the process has been performed, and if so, proceed to the end. If not, the process proceeds to step S74 with the attribute Ax found first with such an attribute and its attribute value Vx.

In step S74, it is checked whether or not the unused attribute Ax appears for the first time, and if so, the process proceeds to step S75, and as an attribute Ax of the diagnosis case other than the new case, belonging to the inspection series (terminal node). , Vx, the first pseudo attribute value not ({V
x}) is given. As the attribute Ax of the diagnostic case belonging to other than the inspection series (terminal node), the second pseudo attribute value * indicating that it is an arbitrary value within the range is given. Then, the process returns to step S73.

If the unused attribute Ax is the already-explained attribute in step S74, the process proceeds to step S76a.

In step S76a, the attribute value Vx is assigned to the attribute Ax in the inspection sequence (terminal node) other than the new diagnosis case.
It is checked whether or not there is another diagnostic case having, and if there is, a return is made to step S73. If not, the process proceeds to step S76b.

In step S76b, it is checked whether or not the inspection series (terminal node) has a diagnosis case having the first pseudo attribute value not (VS) in the attribute Ax other than the new diagnosis case, and if there is, it proceeds to step S76c. Go on, VS to Vx
Is added, and for the diagnosis case having the second pseudo attribute value * in the attribute Ax, the first pseudo attribute value not is added in Ax.
After giving ({Vx}), the process returns to step S73.

If it is determined in step S76b that there is no diagnostic case having the first pseudo attribute value not (VS) in the attribute Ax other than the new diagnostic case in the inspection sequence (terminal node), the process proceeds to step S76d and the attribute Ax is determined. For the diagnosis case having the second pseudo attribute value * in Ax, the first pseudo attribute value not in Ax
After giving ({Vx}), the process returns to step S73.

The generation of the decision tree can be reconfigured by the decision tree generation processing unit 56b. FIG. 8 is a flowchart for explaining the operation of the decision tree generation processing unit 56b in the decision tree reconstructing device 56.

In step S81, the learning data generated by the learning data generation processing unit 56a is read from the pseudo diagnosis case data memory. Next step S8
In 2, the intermediate decision tree is generated by the learning algorithm of the ID3 system using this case set. In step S83, when the decision tree is first created or an intermediate decision tree different from the previously created decision tree, the process proceeds to step S85, and the intermediate decision tree and the learning data are shown as learning data 56c in FIG. Hand over to the pruning section. If it is the same as the previously created decision tree, the intermediate decision tree is output and written in the decision tree data memory.

FIG. 9 shows each division target node L which is used when the decision tree generation processing unit 56b constructs an intermediate decision tree.
Belonging to the class Ck among the learning data belonging to the attribute Ai =
9 is a flowchart for explaining a calculation procedure of the number N (k, i, j) of cases of Vi, j.

In step S91, the sum N (k, i, j) of the weights of the cases belonging to the class Ck and having the attribute Ai = Vi, j is calculated for all the attributes Ai in the node L.
It is checked whether the calculation is completed, and if not, step S9.
Go to 2 and end if completed.

In step S92, a case E in which the attribute Ai has a value Vi, j other than the pseudo attribute and belongs to the class Ck
The sum n (k, i, j) of the weights Wl, n of n is obtained and stored in the number-of-cases memory N (k, i, j). In step S92, the case in which the attribute Ai has the first pseudo attribute not (VS) is checked, and if there is a case in which the contribution calculation to the number-of-cases memory is not calculated, the case En is held and the process proceeds to step S94. If all have been processed, the process returns to step S91.

In step S94, it is checked whether the set of values included in VS and Range (Ai) match, and if they match, N (k, i, N) in step S95.
Wl, n is added to d), and the process returns to step S91. Here, let d be an index corresponding to not (VS). If they do not match, the process proceeds to step S96 and R
Attribute value V included in range (Ai) but not included in VS
For all i and j, N (k, i, j) is given in (B) below.
Calculate with the formula and go to the end.

[0059]

(Equation 3)

FIG. 10 shows that the learning data belonging to each division target node L shown in FIG.
9 is a flowchart for explaining a calculation procedure of weights assigned to each case used in the calculation of the number N (k, i, j) of cases where i = Vi, j, and each case belonging to a split node The procedure of assigning the weight of En to the child node at the end of each arc generated after the node division is shown.

In step S101, it is checked whether or not the weights included in the nodes after the division have been calculated for all the cases included in the division node, and if the calculation has been completed, the process ends.
If not, the process proceeds to step S102.

In step S102, an uncalculated case En
It is checked whether the value Vi, t of the division attribute Ai of (weight Wl, n) is a pseudo attribute value, and if it is not the pseudo attribute value, in step S103, the value of En occupying the node ahead of the arc corresponding to Vi, t The weight is set to Wl, n, and the weight of En occupying the other nodes ahead of the arc is set to 0, and step S101 is performed.
Return to. If it is the pseudo attribute value in step S102, the process proceeds to step S104.

In step S104, the value of the division attribute Ai of the case En is the first pseudo attribute value not (VS), and R
If ange (Ai) = VS, the weight of En occupying the node ahead of the arc corresponding to not (VS) after division by Ai in step S105 is Wl, n, and the node ahead of the other arcs. Setting the weight of En to 0, step S
Return to 101. If not, A in step S106
After division by i, the weight of En occupying the node at the tip of the arc corresponding to Ai = Ai, j is calculated by the following equation (C).
Using (i, j), P (i, j) Wl, n is set, and the process returns to step S101.

[0064]

[Equation 4]

The learning data pruning processor 56c performs pruning on a terminal node basis shown in FIG. 11 and pruning at a non-terminal node shown in FIG. In FIG.
First, in step S111, it is checked whether or not pruning in all the end nodes is completed. If completed, the process proceeds to END, and if not, the process proceeds to step S112. In step S112,
It is checked whether or not there is an arc corresponding to the pseudo attribute value in the check sequence reaching the terminal node that has not been pruned, and if there is, step S11.
After pruning at the non-terminal node shown in FIG.
It proceeds to step S111. If not, the process proceeds to step S113.

In step S113, it is checked whether or not the pruning process inspection has been performed on all the cases at the terminal node, and if completed, the process proceeds to step S111, and if not, the process proceeds to step S114. In step S114, it is checked whether or not the unencumbered case En has a pseudo attribute value in the attribute Ai in the test sequence, and there is another case in which Ai has an actual observation value.
In case 5, the case En is deleted from the pseudo diagnosis case set included in the terminal node, and the process proceeds to step S113. If not,
It proceeds to step S116.

In step S116, it is checked whether or not the unencumbered case En has a pseudo attribute value in the attribute Ai outside the test sequence, and there is another case in which Ai has an actual observation value. In step S117, the case En is deleted from the pseudo diagnosis case set included in the terminal node, and the process proceeds to step S113. If not, step S113 is directly performed.
Go to.

FIG. 12 shows the pruning processing procedure in the non-terminal node of the learning data pruning processing unit 56c. In step S121, it is checked whether or not pruning processing has been performed for all arcs having pseudo attribute values on the check sequence,
If yes, go to end, else step S
Proceeding to 122, the number Nd of cases included in the node after the division regarding the arc which has not been processed yet is obtained.

In the next step S123, the number of cases Na included in the node before the division is calculated, and N is calculated in step S124.
The ratio Nd / Na of d to Na is a predetermined threshold value T
It is checked if it is smaller than h, and if so, all the cases included at the end of the arc corresponding to the pseudo attribute value are deleted, and the process returns to step S121. If not, the process directly returns to step S121.

As described above, in the present embodiment, the decision tree reconstructing device 56 classifies the new diagnosis cases by the diagnosis series defined in the form of the decision tree regarding the possible values of the unexamined part of the diagnosis cases. In the process of obtaining the pseudo observation value, the learning data is reduced by using the intermediate decision tree generated by using the learning data, and the appropriate decision tree for diagnosing the diagnostic cases collected up to that point is reconstructed. Since it is configured, the decision tree can be automatically reconfigured without human intervention. for that reason,
There is no input error in the decision tree optimization data, and the work of decision tree optimization can be reduced.

The embodiments described above are merely illustrative of the present invention and not restrictive, and the present invention can be implemented in various other modified modes and modified modes. Therefore, the scope of the present invention is defined only by the claims and their equivalents.

[0072]

As described in detail above, according to the present invention, the first accumulating means for accumulating decision tree data of a test sequence defined in the form of a decision tree and the second accumulating means for accumulating diagnostic case data. Storage means, a diagnosis means for diagnosing a new case by the decision tree from the first storage means, and if the diagnosis of the new case is wrong at the terminal node, the diagnosis case data from the second storage means is referred to. , If the new case has a new test item, the first pseudo observation that indicates a value other than the value of the corresponding test item of the new case in the value of the corresponding test item of the existing diagnostic case classified in the extracted test series. By giving a value and giving a second pseudo observation value indicating an arbitrary value within the range of the inspection item to the value of the corresponding inspection item of the existing diagnosis case classified outside the inspection series, the unknown observation value Output nonexistent test series as training data Since the learning data generation processing means and the decision tree generation processing means for reconstructing a new decision tree by applying a learning algorithm to the learning data from the learning data generation processing means are provided, the inspection is performed during the diagnosis period. Even when the number of items (attributes) increases, the learning algorithm can be applied to optimize the diagnosis order. as a result,
Since a decision tree can be automatically reconstructed in a short time without human intervention, there is no data input error and the amount of work can be reduced, and a diagnostic result can be obtained in a short time. Can be obtained.

[Brief description of drawings]

FIG. 1 is a block diagram schematically showing a configuration of a conventional diagnostic device.

FIG. 2 is a diagram showing a diagnostic case stored in a diagnostic case data memory.

FIG. 3 is a diagram illustrating pruning of a set of cases in a test sequence of observation values.

FIG. 4 is a diagram illustrating pruning of a set of cases in a test sequence having pseudo observation values.

FIG. 5 is a block diagram schematically showing a configuration of a diagnostic device according to an embodiment of the present invention.

FIG. 6 is a diagram showing an example of a learning data generation processing unit flow in the decision tree reconstruction device.

FIG. 7 is a flowchart for explaining an operation of a pseudo case generation process when the learning data generation processing unit arrives at a terminal node.

FIG. 8 is a flowchart for explaining the operation of the decision tree generation processing unit.

FIG. 9 is a flowchart for explaining the operation of calculating the number of appearances of a class name, attribute name, and attribute value pair used in the decision tree generation processing unit.

FIG. 10 is a flowchart for explaining a procedure for calculating weights of cases before and after node division.

FIG. 11 is a flowchart for explaining a pruning procedure for each terminal node of the learning data pruning unit.

FIG. 12 is a flowchart for explaining a pruning procedure for each non-terminal node of the learning data pruning unit.

[Explanation of symbols]

 50 terminal device 51 decision tree data memory 52 event analysis unit 53 inference control unit 54 diagnostic case data memory 55 history management unit 56 decision tree reconstruction device 56a learning data generation processing unit 56b decision tree generation processing unit 56c learning data pruning processing unit 57 Pseudo diagnosis case data memory

Claims (4)

[Claims] 1. A diagnostic device using diagnostic knowledge in the form of a decision tree, the first accumulating means for accumulating decision tree data of a test sequence defined in the form of a decision tree, and diagnostic case data. The second accumulating means, the diagnosing means for diagnosing the new case by the decision tree from the first accumulating means, and the diagnosis case from the second accumulating means when the diagnosis of the new case is wrong at the terminal node. If the new case has a new inspection item by referring to the data, the value of the applicable inspection item of the existing diagnostic case classified in the extracted inspection series indicates a value other than the value of the applicable inspection item of the new case. By giving a pseudo-observed value of 1 and giving a second pseudo-observed value indicating an arbitrary value within the range of the inspection item to the value of the relevant inspection item of the existing diagnosis case classified outside the inspection series, The test sequence with no observed value of Learning data generation processing means for outputting the learning data, and decision tree generation processing means for reconstructing a new decision tree by applying a learning algorithm to the learning data from the learning data generation processing means, and learning by the generation decision tree A diagnostic device using diagnostic knowledge in a decision tree format, characterized by comprising learning data pruning processing means for reducing data. 2. When the diagnosis item in the node before reaching the terminal node does not have a value corresponding to the observation value of the new case and the diagnosis cannot be made, it is made from a value other than the corresponding observation value. If there is an arc corresponding to the first pseudo observation value, the diagnosis is advanced along the arc, and if not, it is a diagnostic means for creating an arc corresponding to the observation value of the new case and creating a terminal node. Claim 1 characterized by the above.
A diagnostic device according to claim 1. 3. If the learning data generation processing means has a second pseudo observation value in the case of an existing inspection item and the value of the inspection item of the existing diagnosis case classified in the extracted diagnosis series has a second pseudo observation value, If a first pseudo observation value indicating a value other than the corresponding observation value of the case is given and the first pseudo observation value is included, the first pseudo observation value indicating that the corresponding observation value of the new case is also excluded The diagnostic device according to claim 1 or 2, which is a learning data generation processing unit to be given. 4. The learning data pruning processing means, if all the tests up to the terminal node are diagnosed by a test sequence in which all the tests are branched by the observation values, and there is a pseudo observation value in the inspection result, the inspection is performed. If there are other diagnostic cases with observed values in the test sequence, delete the diagnostic cases with pseudo observed values, and if there is a branch with pseudo observed values in the inspection up to the terminal node, that pseudo observed value. If the ratio of the number of cases after division to the number of cases before branching is less than or equal to a threshold value, it is learning data pruning processing means for deleting all the number of cases after the pseudo observation value division. Item 4. The diagnostic device according to any one of items 1 to 3.