KR100231958B1 - Machine failure isolation using qualitative physics - Google Patents

Abstract

The qualitative inference system generates a number of open hypotheses 98 assuming that one or more machine components have failed, and tests each of the hypotheses, using a set of equations for the hypotheses being tested to determine the machine parameters. Mechanical fault isolation is performed by propagating and testing the actual values for Variables associated with the hypotheses of the qualitative physics model are grouped into Type I and Type II variables, and any possible values measured for Type I variables do not match at least one hypothesis, and all are measured for Type II Some possible values will not match at least one hypothesis. The set of sets is considered inconsistent by determining that there is no set of values that can be assigned to all unknown variables in the set resulting in a matching prediction set. Unknown variable values are determined by iteratively setting the variable to each of its possible values and, at each iteration, propagating the unknown variable value and the known variable through the set of models. The model instructor also translates computer workstation user input into usable model information and also provides optimizations that reduce the number of variables and sets in the model.

Description

Machine failure determination method and device

1 is a perspective view of a portable maintenance aid (PMA).

2 shows a user prompt screen.

3 is a flow chart illustrating the overall operation of the fault isolation software.

4 is a data flow diagram illustrating operation of Qualitative Inference System (QRS) software.

5 illustrates a data structure used by the QRS software.

6 is a data flow diagram illustrating the operation of a hypothesis tester in QRS software.

7 is a data flow diagram illustrating operation of a state generator in QRS software.

8 is a flow diagram illustrating steps of a constrained propagator in a state generator.

9 is a flowchart illustrating operation of a core predictor in a state generator.

10 is a data flow diagram illustrating operation of a hypothesis generator in QRS software.

11 is a data flow diagram illustrating the operation of intelligent test selection in QRS software.

12 is a data flow diagram illustrating a model builder.

* Explanation of symbols for main parts of the drawings

30: PMA 32: Display

34: keyboard 36: processing unit

40: User Prompt Screen 42: Query

44, 46 response 92: input process code module

94: practical observation data element 96: hypothesis tester code module

98: Pending Hypothesis Data Elements 99: Preservation Hypothesis Data Elements

100: intelligent tester selection code module 101: component information data element

102: test request data element 103: output processing code module

104: Prediction Tester 105: Hypothesis Generator

106: Model example data element 107: Continuous testing query data element

108: continuous testing response data element 132: state generator

134: Predictive Data Elements 136: Hypothesis Evaluator

142: Constant Explorer 144: First Part Prediction Table Data Element

146: Observation Explorer 148: Second Part Prediction Table Data Element

150: constraint propagator 152: third part prediction table data element

154: core predictor 158: fourth partial prediction table data element

159: Home Tester 232: Hypothesis Controller

234: hypothesis storage data element 236: direct selector

238: set data element 240: variable collector

242: variable data element 244: set cacher

246: Predictive Table Generator 262: Test Classifier

264: Classified test data element 266: Test result gain generator

268: gain data element 270: hypothesis probability generator

272: hypothesis probability data element 274: test result probability generator

276: Estimated Probability Data Element 278: Test Efficiency Generator

280: test efficiency data elements 282: test score generator

284: test score data elements 286: test selector

288 threshold data element 302 user interface

304: Model Organizer 306: Model Component Data File

308: Model Initiator 310: Example Data File

312: Model Tester 314: Tester Case Data File

The present invention relates to the field of computer software, and more particularly to the field of artificial intelligence computer software.

Often, signs of machine failure indicate different interpretations, and it is more cost-effective and time-consuming to exclude any of these interpretations through closer observation of the machine. The process of observing a machine repeatedly and excluding the potential cause of the machine failure is called "failure isolation."

Fault isolation can be done manually with the aid of a failure tree, i.e., a flowchart representation of the repetitive observation / removal steps for fault isolation. Each element of the fault tree requires the user to make a particular observation. A number of branch paths extend from each element, each leading to a different part of the fault tree. The user follows a specific shunt based on the observations required by the open element. At some point in the process, the user will come to an element where the shunt no longer extends, indicating a particular component or group of components that has failed.

For large, complex machines, the failure tree can reach many pages, perhaps several books, making it difficult to observe the failure tree. A solution is to use a program containing information from the fault tree, i.e. a computer with a rule-based fault isolation system. This computer allows the user to make observations and input observation results.

However, in the case of fault trees and rule-based fault isolation systems, all possible failure modes that the user may have to resolve must be determined at the time of its creation. This may not be difficult for simple machines, but may be impossible or at least extremely difficult to implement for more complex machines. It is common for a fault tree designer or rule-based fault isolation system programmer to miss some of the failure modes of a machine. This omission may be due to inattention due to the enormous amount of work, or by intentional decision to keep the size below the actual usage limit.

For a solution to a failure tree or rule-based fault isolation system that cannot isolate all possible failures, see Davis, Randall, “Diagnostic Reasoning Based on Structure and Behavior,” Artificial Intelligence, 24 (1984). , 347-410. Davis proposes a method of fault isolation called so-called "constraint suspension" in which computers generate many models of the machine. Each model assumes various failed components or groups of failed components, and the model most similar to the user's observation indicates that the component or group of components has failed.

The disadvantage of the constraint retention technique is that the modeling of a complex machine with many analog quantities requires a lot of processor intensive work and a lot of system execution time. The solution is described by Hamilton, Thomas P., “HELIX: A Helicopter Diagnostic System Based on Qualitative Physics,” International Journal of Artificial Intelligence in engineering, Vol. 3, No. 3, July, 1988, pp 141-150. Hamilton proposes a modeling technique that represents an analog quantity as a variable that can take a finite set of values, that is, a combination of constraint reservation and qualitative physics. Each of the qualitative finite values represents a range of different analog quantities. However, Hamilton does not provide enough detail for a person skilled in the art to make and use a qualitative physical fault isolation system.

The present invention involves performing separation of mechanical fixation using qualitative physics. The present invention also includes selecting qualitative physics parameters for observation. The invention also includes determining inconsistent hypotheses from a set of hypotheses associated with a qualitative physics model. The present invention also includes predicting values for qualitative physics hypothesis variables and removing inconsistent hypotheses. The invention also includes building a qualitative physics model.

According to the present invention, a number of pending hypotheses are generated, each modeling a machine using qualitative physics, each having a set of confluence describing the behavior of the machine with one or more failed components. Test the hypotheses one at a time by propagating the actual values of the machine parameters to generate predictive sets for the values of the set variables, discarding hypotheses that produce inconsistent prediction sets, and If so, machine component failures are isolated by leaving the hypothesis describing the failed component. In addition, according to the present invention, all possible values of the variable selected for observation will not match at least one of the qualitative physics model hypotheses. Furthermore, according to the present invention, when a matching prediction set is formed for variables that appear in the qualitative model set and are likely to result in the determination or inconsistency of other unknown variables by assigning a predetermined value to the variables in the set (this (If any variable is assigned a predetermined value) or propagates through sets of hypotheses until it is determined that there is no set of values that can be assigned to all unknown variables to produce a matching prediction set. Values are assigned repeatedly. Furthermore, according to the present invention, all possible values for an unknown variable associated with a qualitative physics hypothesis can be eliminated because there is no value that results in a matching prediction set or a single result that the hypothesis can be eliminated or that variable results in a matching prediction set. It is propagated repeatedly through sets of hypotheses to determine whether it can be set to a value. In addition, according to the present invention, a user constructs a model component data file, which is provided as an input to an instantiator, using a graphical user interface, wherein the instantiator is a variable of the model component data file. And other optimizations for reducing the number of sets and generating an example data file. In addition, in accordance with the present invention, example data files and test case data files are provided as inputs to a model tester that executes the model and provides the user with useful information for debugging the model components.

These and other objects, features and advantages of the present invention will become more apparent from the following detailed description of embodiments of the invention as illustrated in the accompanying drawings.

Referring to FIG. 1, Portable Maintenance Aid (PMA) 30 has a display 32, a keyboard 34, and a processing unit 36. PMA 30 is a portable computer, part number A31U18031-3, manufactured by Grumann Electronic Systems Division, Bethpage N.Y., New York. The display 32 is a 7 inch by 9 inch LCD (liquid crystal display). The keyboard 34 is a QWERTY keyboard. Processing unit 36 includes a Sparkstation IE circuit board manufactured by Sun Microsystems Inc., Mountain View, Cal., USA.

PMA 30 is used to perform fault isolation of machines, such as helicopter electromechanical systems. The fault isolation software written in Common Lisp is stored on a hard disk (not shown) located in the processing unit 36. This software allows the PMA 30 to communicate with the user to perform fault isolation of the machine. The software uses the display 32 to prompt the user to observe the machine. The user inputs this observation result through the keyboard 34.

2 shows a user prompt screen 40. The user is given a question 42 and a list of possible responses 44 and 46 to that question. The question 42 used in this example requires a response of yes or no, so the answer that appears on screen 40 is given “YES” 44 or “NO” 46. The user selects one of the responses 44, 46 using the cursor key and then presses the return key to inform the software that the response has been selected. Other user prompt screens may require the user to measure and enter a particular value.

3 is a degree of haze 50 illustrating the operation of fault isolation software, with steps 54 for a rule-based system and steps for a Qualitative Reasoning System (QRS). (58). First, steps 54 for a rule-based system are executed, the implementation and operation of which are already known to those skilled in the art (e.g., Frederick Hayes-Roth, Donald A. Waterman (See Donald A. Waterman, and Building Expert Systems, editors.Addison-Wesley Publishing Company, Inc., Reading Mass. 1983, by Douglas B. Lenat.) Rule-based systems are typically Programmed with information about the faults and their symptoms, they provide a quick answer to any fault determination that the system is programmed to detect, but in the case of complex machines, all of the faults and the associated signs Although predicting possible combinations (and therefore programming all combinations of failures and signs into a rule-based system) is not impossible, at least very difficult to implement. Thus, any particular failure may not be detected by the rule based system, in which case control passes from step 54 of the rule based system to steps 58 of the QRS, Step 58 of can isolate the fault without having to be preprogrammed with all combinations of machine failures and indications.

Execution of the fault isolation software is initiated at a start stage 62 where the rule based system is enabled. In step 63, the user is prompted to enter the observed state. Observation state is a physical description of the state of a particular part of the machine, for example the voltage between two specific points or the particular switch in the "ON" position while the particular indicator is "OFF". After step 63, control proceeds to step 64 where the rule based system attempts to isolate the fault by applying the preprogrammed rule to the observation state.

Some observation states directly indicate a specific fault. For example, you can directly isolate a fault on a dead battery from the observation that the voltage across the battery is zero (assuming there is a rule stating that the battery is exhausted when the measured voltage across the battery is zero). . For example, the observation that the voltage gauge for a battery is read as 0 (assuming proper rules are programmed into the system) indicates that the battery is exhausted, the gauge is malfunctioning, or the wiring between the battery and the gauge is broken.

After step 64, control proceeds to step 65 to determine if the fault has been isolated. At this time, if the failure has been separated for a single component, the process is complete. If the fault has not been separated for a single component, control proceeds to step 66 to determine if further fault isolation is possible. If possible, control returns to step 63 to prompt the user for another observation. Steps 63-66 form a repeating loop, which prompts the user to enter more observations that the software uses for another isolation of the failure.

However, at step 66, it may not be possible to isolate the failure anymore because of the limitations inherent in the rule system (ie, because all possible combinations of observations and associated failures have not been programmed into the system). For example, suppose that the voltage across the measured battery is 12V, the battery voltage gauge is read as 0, and the gauge is observed not to fail. If the rule system is not programmed to take into account the possibility of disconnection between the gauge and the battery, you will encounter difficulties. This observation state does not correspond to any combination of observed conditions and failures that would have been expected if the rule system had been a program. In this case, control proceeds from step 66 to step 67 to disable the rule system. The transition from step 66 to step 67 also corresponds to the transition from step 54 of the rule system to step 58 of the QRS.

After disabling the rule system in step 67, control proceeds to step 68 to enable the QRS. Proceeding to step 69 after step 68, the QRS attempts to isolate the failure. QRS fault isolation will be discussed in more detail later. After performing step 69, it is determined in step 70 whether the fault has been separated. If the fault has been isolated, the process is complete. Otherwise, if the fault has not been separated, the process proceeds to step 71 to determine if further separation is possible. If not possible, the process is complete. In contrast, if possible, proceed to step 72 to prompt the user for observation. Since the transition from steps 54 of the rule system to steps 58 of the QRS is transparent to the user (ie, the occurrence of such a transition is not notified to the user), the prompt the user sees in step 72 is a step This is the same prompt you see in (63). Control proceeds back from step 72 to step 69, where the QRS attempts to isolate the machine failure again.

Unlike rule-based systems, QRS does not directly correlate observations to specific failures. Instead, QRS uses a model of the machine to repeatedly hypothesize a failure of a particular component and, from the user's observations, to the value of various machine parameters such as the amount of current, voltage, and flow rate for each hypothesis. Make predictions about If at any point in this derivation process it is found that the predictions for a particular hypothesis do not match (either with their hypothesis itself or with the observations thereof), the hypothesis is discarded.

QRS represents a machine using qualitative physics computer modeling techniques in which each component of the machine is represented as a black box with multiple terminals and corresponding sets of variables, each variable entering or leaving a terminal (e.g., , Air, fuel, etc.) (eg, flow rate, pressure, temperature, etc.). Each variable may take a finite set value. The behavior of a component is defined by sets, which are sets of qualitative equations that define the relationships between the variables of that component. For example, a pipe can be represented as having two terminals and two variables, one of which is the amount of fluid entering the pipe and the other is the amount of fluid coming out of the pipe. The set describing the behavior of the pipe is indicated to have the same sign as the variable representing the amount of outflow from the pipe. For a more detailed description of qualitative physics, see “The Origin, Form and Logic of Qualitative Physical Laws,” Proceedings of the Eighth International Joint Conference by de Kleer, Johan and Brown See, John Seely. on Artificial Intelligence, Karlsruhe, W. Germany, Aug. 1983 ".

The various components of the machine are grouped hierarchically. A compound component is composed of a plurality of subcomponents, and an elementary component is a component having no substructure. For example, a power supply may be represented as a single component at one level of the hierarchy, but in practice consists of multiple components (eg, capacitors, transformers, etc.) at other lower levels of the hierarchy. Can be. At the highest level of the hierarchy, the entire machine is represented as one composite component. At the lowest level of the hierarchy are all the basic components that make up the machine.

Since the component is modeled as a black box, it may be useful to examine a lower level of the component (ie, go inside the black box to obtain more information) to further isolate the failure. For example, after determining that the power supply has failed, the QRS may deploy the power supply component into subcomponents and continue the fault isolation process at that hierarchical level.

The relationship defined by the various sets of specific components is referred to as "constraints". The QRS calculates the failure effect of a component by suspending (ie, removing) the constraint of that component. For example, for a machine with three components, X, Y, and Z, QRS would leave the qualitative physics model of the machine (i.e., components Y and Z (they fail We will test the hypothesis that component X has failed by creating a model of a machine that includes only a set of machines) that includes only a model of a machine that simulates only). The QRS then uses a user observation and set to generate a prediction set (predicted variable values). If the resulting predictions match, X remains the valid hypothesis, and if they do not match, X is removed as the hypothesis. Hypotheses can also be eliminated if later observations indicate inconsistencies with the predictions. For a detailed discussion of reservations, see Davis Lander, “Diagnostic Reasoning Based on Structure and Behavior,” Artificial Intelligence, 24 (1984), 347-410.

QRS starts with the initial hypothesis that none of the mechanical components has failed. If the predictions generated from the hypothesis match, the QRS software concludes that there are no failures and terminates the fault isolation process. On the other hand, if there is a mismatch, the QRS software generates a number of hypotheses, each corresponding to a failure of a single component of the machine. Predictions associated with each of these hypotheses are generated, and then consistency is tested. If all of the predictions associated with a single component hypothesis are inconsistent (and therefore, the error of all hypotheses is proven), the QRS software generates a new set of hypotheses that correspond to the simultaneous failure of two machine components. If all the dual component predictions do not match, the QRS software generates a hypothesis set that corresponds to three simultaneous failures.

For example, for a machine with three components A, B, and C, QRS first generates a set of sets (ie, all sets for all components) based on the hypothesis that no component has failed. We compute predictions for the hypotheses and test the consistency of these predictions. If the predictions based on the hypothesis match, the hypothesis that no component has failed is verified and the fault isolation process is terminated. If they do not match, the QRS generates three new hypotheses: the hypothesis that only component A failed, the hypothesis that only component B failed, and the hypothesis that only component C failed. The QRS then generates three sets of predictions: predictions for hypothesis A, predictions for hypothesis B, and predictions for hypothesis C. The prediction for hypothesis A is the prediction value that the variable (ie, mechanical parameters such as current, flow, voltage, etc.) will have if hypothesis A is true, and the prediction associated with hypothesis C is the prediction that the variable will have if hypothesis C is true. . Similarly, the prediction associated with hypothesis B is the prediction that the variable will have if hypothesis B is true. If the prediction associated with hypothesis A is inconsistent, hypothesis that A has failed is inconsistent, so hypothesis A is removed from the list of valid hypotheses. On the other hand, if the predictions associated with B and C match, subsequent failure isolation may be required to determine if components B and C have failed.

In the case of the above example, the QRS may remove the hypotheses A, B, and C to conclude that two or more components of the machine failed at the same time. The QRS then assumes that three new hypotheses, namely, hypothesis AB, which assumes components A and B failed simultaneously, hypothesis AC, which assumes components A and C failed simultaneously, and that components B and C were fixed at the same time. Create a hypothesis BC. The QRS then uses these hypotheses to initiate fault isolation. For complex machines, increasing the number of faulty components per hypothesis increases the number of treatments for the computer. When QRS removes all of the N component hypotheses, the user is queried as to whether the QRS software should generate and test N + 1 component hypotheses or terminate all fault isolation. You may not want to wait for testing because QRS generates a new set of hypotheses.

4 is a data flow diagram 90 illustrating the operation of a QRS. The boxes in the data flow diagram 90 represent program modules (ie, QRS software portions) and the cylinders represent data elements (ie, QRS data portions). The arrow between the box and the cylinder shows the data flow direction. Unlike the flow chart, the data flow chart 90 does not show any temporary relationships between the various modules.

The observation by the user is provided to the QRS software as the input signal USERINPUT shown in the data flow chart 90. The USERINPUT signal is processed by the input process code module 92, which converts the key input entered by the user into a format that the QRS software can process. The output of the input process module 92 is stored in the actual observation data element 94, which contains a cumulative history of all observations made by the user after initiation of the fault isolation process. The actual observation data element 94 is initialized with data obtained from the user's observations during the steps of the rule-based system of the fault isolation process.

Hypothesis tester code module 96 uses the actual observation data element 94 and the pending hypothesis data element 98 to generate predictions for potentially valid hypotheses and their respective hypotheses. The open hypothesis data element 98 includes the hypothesis to be tested. Hypothesis tester 96 tests the hypothesis by propagating observation data 94 through sets of hypotheses to generate predictions that should be true for hypotheses that will turn out to be true. During the prediction generation process, hypothesis tester 96 may find inconsistencies. For example, one set subset predicts a positive voltage between two specific points, while another set subset predicts a negative voltage between the same two points. . When this situation occurs, the hypothesis being tested turns out to be an error and is discarded by hypothesis test 96. Hypotheses from pending hypothesis data elements 98 that cannot be discarded by hypothesis tester 96 and predictions associated therewith are output to preservation hypothesis data elements 99.

For example, assume that the pending hypothesis data element 98 includes hypothesis A. Hypothesis tester 96 examines a model of a machine with a bounded component A constraint (i.e., a model that includes all component sets except the set describing the behavior of component A) and predicts the values of the variables. Test the validity of Hypothesis A. While generating hypothesis A predictions, if hypothesis tester 96 finds a mismatch, an error of hypothesis A is found. However, if there is a match, hypothesis A and the prediction associated with hypothesis A are output to conservative hypothesis data element 99.

If the preservation hypothesis data element 99 includes more than one hypothesis, it may be useful for the user to enter more observation states to provide information for removing some of the hypotheses. Intelligent test selection code module 100 is provided with inputs from preservation hypothesis data element 99 and component information data element 101. Component information data element 101 includes experimental data such as component failure rates and test times for potential observations. The intelligent test selection code module 100 uses the predictions from the preservation hypothesis data element 99 and the information from the component information data element 101 to remove the best test for the user to perform (ie, remove one or more hypotheses). At the same time, the observation that is likely to minimize the inconvenience of the user). The intelligent test selection code module 100 outputs the best test information to the test request data element 102. The output process code module 103 is provided with data from the test request data element 102 and converts the test request data element 102 into a human readable format and instructs the user to make the next observation. Provides a signal DISPLAYOUT that causes it to appear on the display 32 of the PMA 30.

Observation by the user may be completely different from the prediction for the hypothesis stored in the conserved hypothesis data element 99. Prediction tester 104 examines the hypotheses in and associated predictions within conservative hypothesis data element 99 and removes hypotheses with predictions that contrast with user observations. For example, a conservative hypothesis data element 99 includes hypothesis A, hypothesis B, and hypothesis C, hypothesis A predicts a positive value for a particular voltage, and hypothesis B predicts a negative value for that voltage. Assume that hypothesis C does not predict the value of that voltage. Further, assume that the intelligent test selection code module 100 selects the voltage measurement to be the best test and, after being prompted, enters the voltage as a positive value. Prediction tester 104 may remove hypothesis B from preservation hypothesis data element 99 because hypothesis B is incorrectly predicted that the voltage is negative, thereby determining hypothesis B's error. Since hypothesis A correctly predicts that voltage is a positive value, hypothesis A remains in the conservation hypothesis data element 99. Prediction tester 104 cannot remove hypothesis C because hypothesis C does not make any predictions about voltage.

The intelligent test selection code module 100 is an observation for the user to make, and it may not be possible to generate all the observations that would remove any hypothesis stored in the conserved hypothesis data element 99. When this situation occurs, the hypothesis generator 105, upon detecting that the test request data element 102 is empty, expands one or more components associated with the hypotheses within the supplemental hypothesis data element 99 to subcomponents. For example, more hypotheses are generated by generating hypotheses based on these subcomponents. Hypothesis generator 105 uses model illustration data element 106, which includes qualitative physics techniques for information about a component and its hierarchical order. Detailed description of the configuration and content of the model example data element 106 will be described later. The output of hypothesis generator 105 is provided to the open hypothesis data element 98.

For example, assume that the conservation hypothesis data element 99 includes a first hypothesis corresponding to a failure of the mechanical power supply and a second hypothesis corresponding to a failure of the mechanical fuel system. Also assume that the intelligent test selection code module 100 cannot provide the user with observations to distinguish between the two hypotheses, so that the test request data element 102 is empty. The hypothesis generator 105 detects that the test request data element 102 is empty, extending the power supply to subcomponents (ie, capacitors, transformers, etc.) and also extending the fuel system to its subcomponents. do. A new set of hypotheses is created based on the subcomponents. The new hypothesis set will be provided by hypothesis generator 105 to the outstanding hypothesis data element 98 (one at a time), tested by hypothesis tester 96, and iterative to prompt the user to observe and remove the hypothesis. Continues. It should be noted that the fault isolation process is completed when the preservation hypothesis data element 99 contains only one hypothesis that assumes the failure of one or more underlying components.

It is also possible for the hypothesis tester 96 to remove all hypotheses stored in the pending hypothesis data element 98 so that the preservation hypothesis data element 99 is empty. This situation occurs when the number of failed components per hypothesis is larger than assumed. Eliminating all hypotheses corresponding to N concurrently failed components indicates that more than N mechanical components have failed. For example, if all hypotheses corresponding to a single component failure are eliminated by hypothesis tester 96 and / or prediction test 104, then logically two or more components failed simultaneously (some component). Note that the likelihood of no failures is initially tested).

When the entire set of hypotheses corresponding to N component failures have been proved error by hypothesis tester 96, hypothesis generator 105 writes the user prompt information to continuous testing query data element 107 to provide N + 1. Ask the user to generate a new set of hypotheses corresponding to the two component failures. The continuous testing query data element 107 is provided as input to the output process module 103, which converts the query 107 into a human readable format and outputs it to the display 32. The user's response to the query is provided by the USERINPUT signal, processed by the input process code module 92, and stored in the continuous testing response data element 108. If the user does not choose continuous testing, fault isolation is terminated. On the other hand, when continuous testing is selected, the continuous testing response data element 108, which is provided as input to the hypothesis generator 105, causes the hypothesis generator 105 to use the new information (using information from the module example data element 106). Generate a hypothesis set, which corresponds to one more failed component than the previous hypothesis set.

The hypothesis generator 105 also provides hypotheses from the conserved hypothesis data element 99 to the open hypothesis data element 98. The hypothesis generator 105 uses the input provided by the test request data element 102 to determine whether the save hypothesis data element 99 includes hypotheses that do not make predictions about the variables requested to be observed by the current user. . Hypotheses that do not predict for the most recently observed variable are sent from an auxiliary hypothesis data element 99 to an open hypothesis data element 98 and tested again by the hypothesis tester 96. Some of the hypotheses tested again may be invalidated as new observations become possible.

5 shows a data structure used by the QRS software. The open hypothesis data structure 110 represents the data stored in the open hypothesis data element 98, which consists of a hypothesis identification element 112B, a prediction table 112C, and a cached set table 112D. (112A). Hypothesis identification element 112B includes information identifying a particular outstanding hypothesis. For example, the hypothesis identification element 112B may identify the pending hypothesis 112A as a hypothesis that assumes a failed power supply.

Prediction table 112C includes all variables (ie, machine parameters) associated with the open hypothesis 112A. The prediction table 112C is configured by the hypothesis generator 105, which determines all the variables of the hypothesis at the time of generation of the hypothesis. Initially, prediction table 112C has no value for the variable. However, each time the hypothesis is tested, the hypothesis tester 96 predicts more values for the variables in the prediction table 112C as more observations become possible.

The cached set table 112D includes all the model sets for the open hypothesis 112A. The cached set table 112D may be indexed by each variable from the prediction table 112C. Elements of the cached set table 112D include all model sets in which index variables appear, thereby providing a hypothesis generator 96 with means to quickly access the set associated with the hypothesis.

The preservation hypothesis data structure 115 represents data stored in the preservation hypothesis data element 99 and includes a first current hypothesis 117A, a second current hypothesis 118A, and an Nth current hypothesis 119A. The first current hypothesis 117A includes a hypothesis identification element 117B and a prediction table 117C. Similarly, second current hypothesis 118A includes hypothesis identification element 118B and prediction table 118C, and Nth current hypothesis 119A includes hypothesis identification element 119B and prediction table 119C. Hypothesis identification elements 117B, 118B, and 119B contain information for identifying a particular stored hypothesis and are similar to hypothesis identification elements 112B of open hypothesis 112A. Prediction table 117C is the same as prediction table 112C from open hypothesis 112A, except hypothesis tester 96 provides values for some of the variables.

6 is a data flow diagram 130 illustrating the operation of the hypothesis tester 96. The state generator 132 is provided with inputs from the actual observation data element 94 and the pending hypothesis data element 98. State generator 132 processes the actual observation 94 and the pending hypothesis 98 to generate predictive data elements 134 that include prediction tables for the particular hypothesis under test. State generator 132 populates the prediction table (from open hypothesis data element 98) with the values of some of the variables computed by processing the set of caches from open hypothesis data element 98 and the actual observation 94. To generate the predictive data element 134. If the state generator 132 detects a mismatch while attempting to predict the value of the predictive data element 134, instead of the predictive table, a null table is stored in the predictive data element 134. The calculation of the actual value is performed by manipulating the LISP expression that represents the set of models. Manipulation of LISP expressions is already known in the art.

Hypothesis evaluator 136 is provided with data from predictive data element 134 and hypotheses from open hypothesis data element 98. For each hypothesis, hypothesis evaluator 136 determines whether prediction data element 134 includes a null table. If not determined, hypothesis evaluator 136 sends the hypothesis (from open hypothesis data element 98) and the associated prediction table (from prediction data element 134) to preservation hypothesis data element 99. . If it is determined to contain, the hypothesis under test is invalidated and not sent.

7 illustrates the operation of the state generator 132 in more detail. The constant finder 142 is provided with a cached set table and a prediction table from the open hypothesis data element 98. The constant finder 142 iterates through the prediction table using a set of caches and populates the variable values of the prediction table with the set defining the variables in constant representation. For example, a battery may be described as a set in which the voltage between its ends is a constant positive voltage. The constant finder 142 outputs a result of filling a constant into an entry of the prediction table as the first partial prediction table data element 144.

The first partial prediction table 144 and the actual observation data element 92 are provided as inputs to the observation explorer 143, which is a first partial prediction table corresponding to the observation entered by the user. Fill in the variable value of 144). For example, if the user measures the voltage across the resistor and inputs the observation, the observation searcher 146 will populate the value of the variable in the first partial prediction table 144 corresponding to the resistor voltage. The observation searcher 146 outputs the result of filling the observation result in the entry of the first partial prediction table 144 as the second partial prediction table data element 148.

The second partial prediction table 148 is provided as an input to the constraint propagator 150, which binds the known variable values (ie, constant searchers ( 142) and the value of the variable already determined by the observation searcher 146 and possibly the variable determined from the previous hypothesis test) and the set of caches from the open hypothesis data element 98 to retrieve more variable values. Determine. Constrained propagator 150 propagates known variables through the set to determine whether one or more sets of each unknown value can be replaced in the form of an unknown variable, such as a constant expression. For example, a set describing the flow of fluid through a valve is such that when the valve is opened, the flow rate exiting the valve is equal to the flow rate entering the valve, and when the valve is closed the flow rate exiting the valve It can be 0. If it is determined that the valve is open (user observation) and the inflow is determined to be a positive value (model constant), then the restraining propagator 150 may determine the flow rate coming from the valve as a positive value. In addition, when the output of the valve is connected to the pipe, the confinement propagator 150 may also determine the inflow and outflow to the pipe.

The variable determined by the confinement propagator 150 then propagates through the confinement propagator along with other known variables. An output third portion prediction table 152 of the constraint propagator is stored. If the constraint propagator detects a mismatch, a null table is stored in the third partial prediction data element 152.

The third partial prediction table 152 is provided as input to the core predictor 154. If the third part prediction table 152 is not a null table, the core predictor 154 iterates through the values for each of the unknown variables of the third part prediction table 152 and, if given a branch variable value, is unknown. Determine if the variable of can have only one possible value. For example, assume that the set for the switch indicates that when the switch is closed, the current coming from the switch is equal to the current entering the switch, and when the switch is open, the current coming from the switch is zero. Also assume that the current coming from the switch has been observed as having a nonzero positive value. Assuming that the switch is open, there is a mismatch (ie, the switch is open at the same time and cannot have a positive current value), so the core prediction indicates that only one possible reasonable value for the state of the switch is closed. The output of the core predictor 154 is stored in the fourth partial prediction table data element 158. It should be noted that core predictor 154 may find inconsistencies for certain hypotheses. For example, using the example of the above switch, suppose that the input current to the switch was observed to be negative. As long as the output current of the switch is positive regardless of opening or closing of the switch, it is a mismatch that the input voltage of the switch is negative. Thus, the hypothesis under test is invalid. That is, the third partial prediction table 152 that is being processed by the hypothesis and state generator 132 associated with the set of caches turns out to be false. When this situation occurs, the core predictor 154 nulls (i.e. sets null) the fourth partial prediction table 158.

The fourth partial prediction table 158 is provided as input to the hypothesis tester 159. The hypothesis tester 159 is given a prediction (i.e., a variable having a value set as the prediction) included in the fourth partial prediction table 158, and at least one combination value is added to the remaining unknown variables that do not cause any discrepancies. Determine if it can be allocated. Of course, if the fourth partial prediction table 158 is a null table, the hypothesis tester 159 sends the null table to the prediction data element 134 and the hypothesis tester 136 may discard the hypothesis.

However, if the fourth partial prediction table 158 is not a null table, the hypothesis tester 159 assumes values for each of the unknown variables, and then determines whether a matching set of predictions can be derived from these values. Determine. The software stores the values for the unknown variables in the fourth part prediction table 158 (in this way, temporarily converting the unknown variables to known variables), propagating all known variables, and then transferring itself. Use a calling routine. During propagation, if an inconsistent prediction set occurs, hypothesis tester 159 goes back through its recursion routine to assume various values of the variable. If a matching set of values for all unknown variables is found, hypothetical tester 159 predicts fourth-part prediction table 158 (recovering unknown variables to their original state) with predictive data elements 134. Send to. If no matching set is found, hypothesis tester 159 provides a null table to predictive data element 134.

Although hypothetical tester 159 can randomly assign values to unknown variables, this is very inefficient in terms of processor time, because the number of random combinations is huge. For example, if there are 20 unknown variables, each with three possible values, the number of random combinations is more than three billion. Thus, instead of randomly assigning a value to an unknown variable, the hypothesis tester 159 assigns a value to the variable using a dynamic hypothesis ordering process.

The dynamic assumption ordering process is the process of locating a target set, assigning values to the variables that appear in the maximum number of target sets, and propagating the variable assignments. A target set is a set that assigns a value to one of the set variables so that the value of another unknown variable can be determined or an inconsistency can occur and the hypothesis can be rejected. In the simplest example, the target set is a set that describes that variable V1 is equal to variable V2. When one value is assigned to the variable V1, a value for the variable V2 may be determined. It is also possible for a mismatch to occur for every assignment of a value to V1 to enable the hypothetical tester 159 to send a null table to the prediction table data element 134. For example, the first set states that variable V1 is equal to variable V2, the second set states that variable V1 is equal to -V2, and the third set states that variable V1 is (variable V2 + positive constant). Assume that we describe the same as). There are no combinations of variables that can solve the constraints on V1 and V2 by the three sets. If the hypothetical tester 159 initially selects V1 or V2 for the assignment of values rather than randomly selecting a variable that does not appear in the target set, an inconsistency will soon be found.

8 is a flow chart 180 illustrating a more detailed operation of the constrained propagator 150. In a first step 182, iteration through known variables in the second partial prediction table 148 is controlled. In step 182, the iteration counter is initialized first and then incremented for each subsequent execution. The remaining steps of flow chart 180 operate on one variable at a time. When the iteration counter reaches the end of the known variable list of the second partial prediction table 148, execution is complete. If not, proceed to step 183, where all sets associated with the variables (from the cached set table) are examined and, if possible, using all known variable values from the second partial prediction table 148. Is solved. If the unknown variable is represented only as a known variable (ie all variables except one of the variables in the set are known), then the unknown variable value may be determined at step 183. Control proceeds from step 183 to step 184 to determine if a mismatch is found. For example, if two conflicting predictions are made for the same variable), for example, one subset of the set predicts that a particular variable is positive, while the other subset of the set predicts that the variable is negative. ) Causes inconsistencies. If an inconsistency is found in step 184, control passes to step 185, where the third partial prediction table 152 is null and execution of the constraint propagator 150 ends.

If no inconsistency is found in step 184, control proceeds from step 184 to step 187 to determine if any unknown variable value was found in step 183. If a new known variable is found, control proceeds from step 187 to step 188, where the new variable is added to the known variable list. Control returns from step 188 to step 182, and the iteration counter is incremented. If no new variable was found in step 183, control returns from step 187 to step 182.

9 is a flowchart 190 illustrating a more detailed operation of the core predictor 154. In a first step 192, iteration through the unknown variable of the third partial prediction table 152 is controlled. In step 192, the iteration counter is initialized first and then incremented for each subsequent execution. The remaining steps of flow chart 190 operate on one unknown variable at a time. If the process is not complete in step 192, control proceeds from step 192 to step 193 to check the active variable for one or more valid values. Since qualitative physics is used to model the machine, all variables, including those representing actual analog quantities, have a finite number of possible values. In step 193, the variable is repeatedly set to all possible values that the variable can take. In step 193, the variable has not yet been set to all possible values, and control proceeds from step 193 to step 194, propagating the variable value through the confinement propagator. Step 194 is similar to constraint propagation illustrated by flow diagram 180. Control proceeds from step 194 to step 195 to determine if propagating hypothesized values for unknown variables through the set resulted in inconsistencies. If it has resulted in a mismatch, control returns from step 195 to step 193 and makes another iteration for that variable (ie, another value is selected for that variable). If propagating the variable values through the set did not result in inconsistency, control proceeds from step 195 to step 196, where the value is added to the list of possible values for that variable. Control then returns from step 196 to step 193 for another value for the variable to initiate the test.

After all possible values for the variable have propagated through the set, control proceeds from step 193 to step 197 to determine if any of the variables resulted in the prediction set with matching prediction values. If no value exists for the variable that would result in a matching set of predictions, control proceeds from step 197 to step 198, where the prediction table is null and execution ends. This hypothesis is not true because one of the variables cannot have a value that produces a matching prediction set. If the predicted value is not zero, control proceeds from step 197 to step 199 to determine if there is only one value resulting in a matching set of predictions for that variable. If so, then control passes from step 199 to step 200 to add the variable and value to the fourth partial prediction table 158. If only one value of the variable results in a matching set of predictions, the variable must be equal to the value for the hypothesis under test that is being determined to be true. Control returns from step 200 and step 199 to iteration step 192 to test the next unknown variable.

10 is a data flow diagram 230 illustrating the operation of the hypothesis generator 105. The hypothesis generator 105 generates a hypothesis through a number of methods as follows. The hypothesis generator 105 may generate a new hypothesis by extending the components associated with the current hypothesis from the conservation hypothesis data element 99. The hypothesis generator 105 may generate a new hypothesis by assuming the maximum number of simultaneous component failures. In addition, hypothesis generator 105 may communicate the current hypothesis and association predictions from conservative hypothesis data element 99 to the pending hypothesis data element 98.

Data from the test request data element 102 is provided to the hypothesis controller 232, which, when the test request data element 102 detects that the test request data element 102 is empty, receives the data from the model example data element 106. The information is used to extend components associated with the hypothesis from the conservation hypothesis data element 99 to subcomponents so that a new hypothesis is generated. For example, if the conservative hypothesis data element 99 includes a single hypothesis that assumes a failure of the mechanical power supply, then the hypothesis controller 232 may include sub-components of the power supply (eg, capacitors, transformers). A number of hypotheses will be generated that correspond to failures of the formwork, bridge rectifier, etc.). Hypothesis controller 232 can determine the subcomponents of the component, because the model example data element 106 includes a data structure that identifies the subcomponents of each composite component.

If the preservation hypothesis data element 99 is empty, the hypothesis controller 232 writes data to the continuous testing query data element 109 so that the user has a hypothesis set with one more failed component than the previous hypothesis set. Determine whether you want to continue troubleshooting. The user's response is provided to the continuous testing response data element 108, which is provided as input to the hypothesis controller 232, which hypothesis controller 232 uses this response. Determine whether the hypothesis will occur continuously.

If the test request data element 102 is not empty (i.e., the intelligent test selection code module 100 prompts the user to observe), the hypothesis controller 232 may draw the hypothesis from the preservation hypothesis data element 99. Send to open hypothesis data element 98 to allow hypothesis tester 96 to perform another test. A hypothesis that predicts the value for a variable that is prompted to be observed by the user (ie, a test stored in test request data element 102) is not passed because another test predicts a new value for that hypothesis or The hypothesis cannot be eliminated. For example, suppose that the conservation hypothesis data element 99 includes hypothesis A, which predicts positive fluid flow through a specific conduit, and hypothesis B, which does not predict fluid flow through the conduit. If the test request data element 102 includes a prompt to allow the user to observe the fluid flow through the conduit, then the hypothesis controller 232 will hypothesize B (since hypothesis B makes no predictions about the fluid flow). Is passed to open hypothesis data element 98, but will not pass hypothesis A to open hypothesis data element 98 (since hypothesis A predicts positive fluid flow through the conduit). It should be noted that if the user observes a negative or zero fluid flow through the conduit, the prediction test 104 will remove the hypothesis A from the conservation hypothesis data element 99.

The hypothesis controller 232 stores the hypothesis (the hypothesis from the newly generated hypothesis or preservation hypothesis data element 99) in the hypothesis storage data element 234. Hypothesis storage data elements 234 are provided as inputs to the set selector 236, each of which is stored in the hypothesis storage data elements 234 using data from the model example data element 106. Determine a set of models for the hypothesis of. Set selector 236 stores the set in set data element 238.

Set data element 238 is provided as input to variable collector 240, which determines a unique variable for each set of sets and stores its output in variable data element 242. The variable data element 242 and the aggregate data element 238 are provided as inputs to the aggregate cacher 244, which generates a cached aggregate table, which aggregates the respective aggregate table. Indexed by variable, each element of the table contains all the sets in which the index variables appear (eg, variable V1 appears in sets C1, C5, and C6, and variable V2 appears in sets C2, C5, etc.). The cached set table is used by hypothesis tester 96 to test the hypothesis without having a search set for variable generation.

For the newly generated hypothesis, a variable data element 242 is provided as input to the prediction table generator 246 which generates an empty prediction table. For hypotheses obtained from the conserved hypothesis data element 99, an existing prediction table (with some variable value already determined) is used. The output of the hypothesis generator 110 is recorded in the open hypothesis data element 98, which is a set of hypotheses (one at a time), with the associated cached set table for each of the hypotheses, and (in the case of newly generated hypotheses, any values). The relevant prediction table for each hypothesis.

11 is a data flow diagram 260 illustrating the detailed operation of the intelligent test selection core module 100. The output from the retention hypothesis data element 99 is provided to the test classifier 262. The test classifier 262 examines the predictions associated with each hypothesis from the conservation hypothesis data element 99 and classifies each variable as a type I test, a type II test, or a type III test, and the type I test is a conservation hypothesis. Is a user-permissible observation that ensures that at least one hypothesis from data element 99 can be ignored, a type II test is an observation that may or may not be ignored, and a type III test is a hypothesis. This is an observation that cannot be ignored. The output of the test classifier 262 is stored in the classified test data element 264.

As an example of test classification, the hypothesis that the conservation hypothesis data element 99 includes hypothesis A and hypothesis B, where hypothesis A predicts that a specific current will be greater than or equal to zero, and hypothesis B predicts that the current will be less than zero lets do it. The test classifier 262 will consider the current as a type I test because, regardless of the actual value of the current, a user observes (and inputs to the QRS) that the current ensures the elimination of hypothesis A or B. Because it has. Continuing with this example, assume that hypothesis A predicts a voltage above zero at a point and hypothesis B predicts a voltage below zero at that point. The test classifier 262 will consider the voltage measurement as a type II test because the user may not be able to remove hypothesis A or hypothesis B by observing the voltage. If the user measures a voltage of zero, none of hypothesis A or hypothesis B can be removed, but if the user measures a non-zero voltage, none of hypothesis A or hypothesis B can be removed. Also assume that both hypothesis A and hypothesis B do not make predictions for a particular fluid flow. The test classifier 262 will then consider the fluid flow as a type III test.

An output of the retention hypothesis data element 99 is provided to a test result gain generator 266, which, for each possible value of each variable, hypothesis from the retention hypothesis data element 99. If the variable is equal to a certain value, the part to be discarded is determined. For example, the conservation hypothesis data element 99 includes ten hypotheses, three of which predict that a certain current will be positive or zero, and four predict that the current will be negative. Assume that the remaining three do not make any predictions about the current. The gain for positive current will be 4/10 and the gain for negative current will be 3/10. Test result gain generator 266 is stored in gain data element 268.

The outputs of the stored hypothesis data element 99 and the component information data element 101 are provided to a hypothesis probability generator 270, which is an experimental component from the component information data element 101. Failure rate information is used to predict the probabilistic effectiveness of each hypothesis from the conserved hypothesis data element 99. The output from hypothesis probability generator 270 is stored in hypothesis probability data element 275.

The hypothesis probability data element 272 and the retention hypothesis data element 99 are provided as inputs to the test result probability generator 274. Test result probability generator 274 predicts the expected value that the user will observe for each variable. The value of a variable predicted by a high probability hypothesis is more likely to be observed than the value of a variable predicted by a low probability hypothesis. For example, assume that the conservation hypothesis data element 99 includes hypothesis A, which predicts that a particular fluid flow will be zero, and hypothesis B, which predicts that fluid flow will not be zero. It is also assumed that hypothesis A is considered to have an 80 percent probability by probability generator 270 and hypothesis B is considered to have a 20 percent probability by hypothesis probability data element 270. At this point, the test result probability generator 234 will determine that the user is 80 percent likely to observe fluid flow as zero and the user is 20 percent likely to observe fluid flow as nonzero.

The output from the test result probability generator 274 is stored in the predicted probability data element 276, which, together with the gain data element 268, serves as input to the test efficiency generator 278. Is provided. For each variable, the test efficiency generator 278 determines the efficiency with which the user measures the variable by calculating the sum of the product of the expected probability and the gain for each value that the variable may have. For example, suppose the variable X has three possible values: minus, zero, plus. It is also assumed that the gain for measuring the variable negatively is 1/10, the gain for measuring the variable as 0 is 2/10, and the gain for measuring the variable as positive is 6/10. Also assume that the probability that X is negative is 25 percent, the probability that X is zero is 70 percent, and the probability that X is positive is 5 percent. At this time, the efficiency of measuring the variable X is determined by the following equation.

Efficiency of X = (0.10 × 0.25) + (0.20 × 0.70) + (0.60 × 0.05)

The output of the test efficiency generator 278 is stored in the test efficiency data element 280, and the test efficiency data element 280, along with data from the component information data element 101, for the test score generator 282. Provided as input. Test score generator 282 divides the test efficiency for each variable by the test time for each variable (from component information data element 101) for each variable stored in test score data element 284. Provide test scores. For each variable, test score generator 282 determines a desire prompting the user to observe the machine parameter represented by that variable. For two variables with the same efficiency, the longer the user measures, the lower the test score will be. In addition, some variables, such as the internal frictional force of the mechanical parts, may render the user impossible to measure, allotting test time to infinity. For details on the theory and calculation of efficiency, see Von Neumann John and Theory of Games and Economic Behavior by Morgenstern, Oskar, Princeton, princeton University Press, 3rd edition (1953). .

The test score data element 284 and the classified test data element 364 are provided as inputs to the test selector 286, which is intended to determine the best observation that a user makes. The test selector 286 is also provided with a third input from the threshold data element 288, which includes a threshold for each test type (in this embodiment, the type The threshold for I is 0.5 and the threshold for Type II is 100). The test selector 286 selects the type I test with the highest test score (ie, user observations that remove at least one hypothesis from the retention hypothesis data element 99). However, if the highest scoring Type I test has a score lower than the threshold for the Type II test, the test selector 286 selects the highest scoring Type I test with a test score higher than the threshold for the Type II test. If there is a type I, type II test with a test score higher than the respective threshold, the test selector 286 does not select the test. The output of the test selector 286 is written to the test request data element 102. If test selector 286 writes nothing to test request data element 102, hypothesis generator 105 detects that test data element 102 is empty to extend the hypothesis from preservation hypothesis data element 99. To start things.

12 is a data flow diagram 300 illustrating a model builder, which generates model example data elements 106 for use by the QRS software. The model example data element 106 is off-line by running a model builder on a computer workstation, such as Symbolics 3640 by Symbolic Inc. of Burlington, MA. It is then delivered to the PMA 30 to be part of the QRS software.

Input to the model builder is via a graphical user interface 302, which is described in “HELIX: A Helicopter Diagnostic System Based on Qualitative Physics.” Hamiton, Thomas P., International Journal of Artificial Intelligence in Engineering, Vol. 3, No. 3, July 1988, pp 141-150. User input from the graphical user interface 302 is provided to the model configurator 304, which processes the user input and stores the model component data file 306 to be stored on the workstation's disk. Occurs. The model component data file 306 contains the definitions of the base model components (ie terminals, variables, and sets of the base components) and the definitions of the composite components, which you can use to interconnect the base components or It can be created by interconnecting other composite components. The data stored in the model component data file 306 takes the form of a LISP expression having a configuration known to those skilled in the art. The interconnection of the components of the component data file 306 defines a model hierarchy such that there is a base component at the lowest level of the hierarchy and a single composite component representing the machine being modeled at the highest level.

The component data file 306 is provided as input to the model insulator 398, which takes into account the characteristics of the model components, parameters, and connections, The component data file 306 is converted into a format optimized for processing by the QRS software of the PMA 30. The conversion result by the model instructor 308 is output to the example data file 310, which is stored on the workstation's disk and passed to the QRS software on the PMA to provide the model example data element. (106). The conversion performed by the insulator 308 is a set conversion from infix notation to prefix notation, keyword extraction from conditional sets, and a dictionary of variables used in each set for quick access. Sorting, and conversion of data variable types from LISP attribute lists to LISP symbols.

Instructor 308 also reduces the number of variables and constraints in the model by reducing constraints. Constraint reduction examines a set of components, in the form V1 = V2 or V1 = -V2, and terminal variables (that describe the condition of the terminal of the component) are simple sets that do not exceed one of the variables. It includes removing it. Other variables may be terminal or non-terminal variables. One non-terminal variable and a simple set are removed by replacing the non-terminal variable in all the sets for the component with another variable (or the negative value of the other variable described above in the case of a simple set of the form V1 = -V2). . One limitation is that the possible qualitative values for non-terminal variables should be a subset of the possible qualitative values for other variables.

The constraint reduction process begins at the optimal level of the model hierarchy. Variables removed at the lowest level are also removed from higher hierarchical levels. At subsequent levels of the hierarchy, more variables may be removed, because variables describing one level of terminal may not describe the terminal of the component at a higher model hierarchy level. For example, when the base components are grouped together, many of the base component terminal variables of the power supply become non-terminal variables.

For the QRS to work properly, the example data file 310 must contain an accurate representation of the system being modeled. The example data file 310 is provided as input to the model tester 312, which interacts with the user through the graphical user interface 302 to execute the components of the example data file 310. do. The model tester 312 utilizes qualitative physics to execute the component to detect and diagnose failures or generate operating states of the component, providing information to the user through the graphical user interface 302. For example, for a valve in the example data file 310, component tester 312 may be configured to provide a first state in which the valve is closed such that the flow rate from the valve and the flow rate into the valve are zero, the valve is open. A second state in which the flow rate entering the valve is positive and the flow rate exiting the valve is positive, and the valve is opened so that the flow rate entering the valve is negative and the flow rate exiting the valve becomes negative And a fourth state in which the valve is open and the flow rate entering the valve is zero and the flow rate exiting the valve is also zero.

The occurrence of a state allows the user to debug the component model. If the user provided too many constraints on a component, the model tester 312 will not generate all possible states for that component. For example, using the valve example described above, if the user provided too much restraint for the valve, the model tester 312 may only occur in three of the four operational states described above. Likewise, if the user provided too little restraint, the model tester 312 may generate an excessive and illegal condition. For example, using the valve example described above, if the user did not provide a restraint of the valve specifying that the outflow and inflow for the valve were zero when the valve was closed, the model tester 312 would close the valve and It may also occur that the inflows and outflows for these are non-zero positive values.

The user may also optionally generate a test case data file 314, which is stored on the workstation's disk. The test case data file 314 includes a predetermined set of assigned variable values of the components stored in the example data file 310. For example, test case data file 314 may include first and second sets of assigned variable values for electrical resistance. In the first set, the current flowing through the resistor has a positive value, and the voltage across the resistor has a positive value. In the second set, the current flowing through the resistor is negative and the voltage across the resistor is also negative. Test case data file 314 need not include all test cases for a particular component. In the case of the resistor example described above, the case where the current and voltage for the resistor is zero is not used.

The test data file 314 may also include a set of variable values that violate invalid cases, ie, a set of components. For example, the test case data file 314 may include a case for a resistor in which the voltage across the resistor has a positive value and the current flowing through the resistor has a negative value. Test case data file 314 is provided as input to component tester 312, which replaces values from a predetermined set of cases with variables of the component under test. Component tester 312 reports the results to user via user interface 302. As the string at which each instance in the test data file 314 is provided by the user at the time of creation is also reported to the user via the user interface 302, for example the string for the invalid case is It can identify itself.

Although the QRS software is illustrated as for the PMA 30 and the model configuration is set up on a Symbolics 3640 workstation, those skilled in the art will appreciate that the QRS or model configuration can be applied to various computational systems. Similarly, while QRS software is written in LISP, the present invention may be implemented using any computer language capable of supporting the required functionality. Although the fault isolation system illustrated herein uses QRS after using a rule based system, it will be understood by those skilled in the art that the QRS can be operated without using a rule based system at all.

In addition, although qualitative physics is used for fault isolation and debugging of model components, it will be understood by those skilled in the art that qualitative physics may be used for many other applications other than the applications illustrated herein. The invention can also be used for fault isolation for any type of system that can be modeled qualitatively. In addition, features of the present invention regarding improvements to qualitative physics modeling (such as core prediction, dynamic hypothesis ordering processes, and aggregation cache processes) have applications other than those illustrated herein. The constraint propagation feature of the core prediction and dynamic hypothesis ordering process can be used to invalidate the mismatch model (without generating a mismatch prediction set whenever a variable cannot take any possible value). Similarly, even if the resulting fault isolation system is performed in a degraded manner, qualitative physics modeling can be used for fault isolation without using core prediction, dynamic hypothesis order processors or aggregate cache processes. Although the dynamic assumption ordering process is illustrated herein as selecting variables that appear in the maximum number of target sets for assignment of values, other approaches such as selecting variables that appear in the second or third maximum number of target sets. We can still derive some of the advantages of the dynamic assumption ordering process.

The thresholds illustrated herein for Type I and Type II of the intelligent test selection code module 100 may be changed. Likewise, the specific equations used in the calculation of the test scores, or the criteria used herein, may be modified without departing from the spirit and scope of the invention. The intelligent test selection code module 100 illustrated herein, in addition to fault isolation, determines the best measurements for testing of the component when designing the component, and thus places test points for the component. It has another application.

The order of assumptions about the number of concurrently occurring component failures illustrated herein (eg, no component failure, one component failure, simultaneous failure of two components, etc.), and The order of hypothesis generation may be modified without departing from the spirit and scope of the invention. Similarly, prompting the user and determining the response each time the number of components assumed to be concurrent failures is increased may be eliminated by the QRS automatically increasing the number and continuing failure isolation without notifying the user. The QRS may operate on a model consisting entirely of basic components, and thus do not require that the model included in the model example data element 106 be hierarchical.

Although one application to the model builder (ie, using the model builder to construct the model example data element 106 of QRS) is illustrated herein, those skilled in the art can have many other applications. Will know. Processes other than those illustrated herein can be used in the generation of model example data elements 106. In addition, QRS models only the subset of optimizations exemplified herein (eg, constraint reduction, keyword extraction, conversion of data variable types from LISP attribute lists to LISP symbols, etc.) by the instance 308. It may operate on a model example performed for, but as a result may degrade the performance of the QRS. The QRS may also be configured to operate directly on the component data file 306, but such a system may be slower to execute than the embodiments illustrated herein. Model examples may be tested using methods other than those illustrated herein.

In the present invention, the user observes, but provides a data connection between the machine in which the fault isolation is being performed and the PMA 30, through which the QRS requests information from the machine, and the machine also connects to the data. It is possible to automate the observation acquisition process by providing the information through. In addition, QRS does not prompt additional observation users if the observation data element 94 contains enough information to isolate the failure for a single base component upon initialization by observation from a rule-based system. Fault isolation can also be performed without Also, by using other methods (including random selection) to determine what observations to perform, fault isolation can be performed without the intelligent test selection code module 100, but in this case QRS is required to isolate the machine fault. It may take longer.

While the invention has been shown and described with respect to exemplary embodiments, those skilled in the art will recognize that various changes, omissions and additions may be made without departing from the spirit and scope of the invention.

Claims (25)

How to isolate a failure of a machine by testing a number of machine models—each of the models assumes that a number of machine components have failed, and the test propagates the actual values of machine parameters through the equations of the model Generating a set of predictions for a value of a machine parameter, discarding an equation of the model that produces an inconsistent prediction set in the model, and matching in the model Storing an equation of a model that produces a consistent set of predictions, wherein the model is generated using qualitative physics descriptions of the mechanical components, and machine parameters. Obtaining an additional actual value of and obtaining a value for the variable that does not match the additional actual value. Discarding each predictive model. The method of claim 1, wherein obtaining additional actual values of the machine parameters comprises: placing a query on a computer display to prompt a user for a particular input; and processing a user keyboard input to respond to a user's response. Determining the failure of the machine. The method of claim 1, wherein the generating step assumes that the first model set has not failed any component, the second model set assumes that one component has failed, and the third model set has two Sequentially ordering the generation of the model, assuming that the two components have failed at the same time, and every subsequent model set assumes that one more component has failed at the same time than each of its preceding model sets. Machine failure separation method comprising a. 4. The method of claim 3, wherein ordering the generation of the subsequent model set occurs only as a user responds positively to the query for permission to generate the subsequent model set. 5. The method of claim 4 wherein the qualitative physics technique is arranged hierarchically so that some components are made up of subcomponents. 6. The machine of claim 5, wherein a new model is created that assumes a failure of a subcomponent of a failed component associated with the stored model, as one or more of the stored model cannot be discarded by prompting a user for input. How to isolate faults. The method of claim 1, further comprising: forming a type I machine parameter set, for each of the type I machine parameter sets, all possible values of the parameter cause the prediction associated with at least one of the models to be inconsistent. And selecting a parameter for observation from said type I machine parameter set. 8. The method of claim 7, further comprising calculating a test score for each value of the type I machine parameter set, wherein the test score indicates a desirability of observation for each of the parameters. The selecting step includes selecting a particular parameter of the parameter for observation from the type I set having the highest test score. 10. The method of claim 8, further comprising: forming a type II parameter set—for each of the type II parameter sets, not all but some possible values of the type II parameters match the prediction associated with at least one of the models. And calculating a test score for each of said type II parameter sets, said test scores indicating a desired degree of observation for each of said parameters from said type II parameters. The selecting step includes selecting a particular one of the parameters for observation from the type II parameter set having the highest test score if the test scores for all parameters of the type I set are lower than a predetermined threshold. How to isolate machine failure. The method of claim 9, wherein the test scores for the Type I and Type II sets are as follows: calculating a gain for each possible value of each parameter, wherein the parameter is equal to the value. Calculate the number of discarded models, if observed, by dividing the total number of models-and correlating the assumed component failure and experimental component failure rate information for each model. Calculating the probability of validity for each model, and the probability of measuring each possible value for each parameter by summing the estimated validity for each model predicting a particular value for the parameter. Calculating a value, and a probability of measuring each value for each parameter and a gain for each value of each parameter. Calculating the observed availability of each of the parameters by summing the product by parameter, and dividing the test utilization of each parameter by the amount of stored, experimentally derived time needed to observe each parameter. A method for isolating a machine failure, which is calculated by calculating a test score for a machine. 11. The method of claim 10 wherein the predetermined threshold is 0.5. A method of determining inconsistencies of a qualitative physics model, wherein in the case of a first-stage selection of a particular parameter from a parameter table associated with the model, other parameters may be derived or the model may be considered inconsistent. The particular parameter appears in an equation for a model that includes a parameter that enables the parameter; and, after the first selection step, a second step of temporarily assigning a constant value to the parameter, wherein the constant value is assigned to the parameter. Is a value included in the list of possible values, and, after the second step, a third step of propagating the constant value through an equation of the model to determine a match, and a specific constant value of the constant value matches. Until all the possible values for the parameter result in a set of predictions A fourth step of repeatedly performing the second and third steps until assigned, and if the third step results in a matching prediction set by assigning a specific constant value of the constant value to the parameter, the constant value Is propagated through an equation of the model to determine temporary constant values for other unknown parameters, and subsequent to the fourth and fifth steps, the first, second, third, Using a recursion to perform steps 4, and 5 repeatedly—the recursion results in a new value each time the fourth step results in all possible values assigned to the parameter without causing a matching prediction set. Return to assign the previously assigned parameter, and generate a value that can be temporarily assigned to an unknown parameter of the model, generating a matching set of predictions for the model. If the sum does not, the model is deemed inconsistent-way mismatch determination of a qualitative physics model comprising a. A method of predicting a value for an unknown parameter associated with a qualitative physics model, comprising: a first step of temporarily assigning a constant value to the parameter, the constant value being included in a list of possible values for the parameter; Subsequently, a second step of propagating the constant value through an equation of a model to detect inconsistency, and a third step of continuously performing the first and second steps for all possible values for the parameter; Subsequent to the third step, if only one specific constant value for the parameter results in a matching prediction set for the model, assigning the constant value to the parameter permanently. Predicting a value for an unknown parameter associated with the. A method of determining a discrepancy of a qualitative physics model, the method comprising: a first step of temporarily assigning a constant value to the parameter, the constant value being included in a list of possible values for a parameter of the model, and subsequent to the first selection step, A second step of propagating a constant value through an equation of the model to detect inconsistencies, a third step of repeatedly performing the first and second steps for all possible values for the parameter, and If the constant value does not result in a matching set of predictions for the model, then the model is considered inconsistent. A method of determining a discrepancy of a qualitative physics model to predict a value for an unknown parameter associated with the qualitative physics model, the method comprising temporarily assigning a constant value to the parameter, the constant value being included in the list of possible values for the parameter. A first step, subsequent to the first selection step, a second step of propagating the constant value through an equation of a model to detect inconsistencies; and the first and second steps for all possible values for the parameter. A third step of repeatedly performing and subsequent to the third step, if only one specific constant value for the parameter results in a matching set of predictions for the model, permanently assigning the constant value to the parameter Assigning, and following the third step, any constant values for the parameters that match the predictive predictions for the model. If not causing said model, said model is considered to be inconsistent, the method for predicting a value for an unknown parameter associated with a qualitative physics model. The method of claim 1, wherein the generating step comprises: locating an equation that provides that a first parameter that is not associated with a terminal of a component of the model is equal to a second parameter or a negative second parameter—the The possible value of the second parameter is a superset of the possible values of the first parameter—the step of eliminating the equation and the first parameter occurring in the remainder of the equation, the second parameter or the negative A method of isolation of a mechanical failure comprising the step of replacing with a second parameter. A device for isolating the failure of a machine has a means for testing a number of machine models, each of which assumes that a number of machine components have each failed, and the test means is adapted to determine the machine parameters through the equations of the model. Means for propagating an actual value to generate a prediction set for the value of a machine parameter, means for discarding an equation of the model that produces an inconsistent prediction set in the model, and means for storing a matching prediction set in the model. Means for generating the model using qualitative physics techniques of the machine components, means for obtaining additional actual values of machine parameters, and for variables that do not match the additional actual values. Means for discarding each model for predicting a value for the failure. 18. The apparatus of claim 17, wherein the means for obtaining additional actual values of the machine parameter comprises a computer display for prompting the user for a particular input and a user keyboard for receiving a response from the user. 18. The apparatus of claim 17, further comprising: means for forming a type I machine parameter set, for each of the I machine parameter sets, all possible values of the parameter cause the prediction associated with at least one of the models to be inconsistent; And means for selecting a parameter for observation from said type I machine parameter set. In an apparatus for determining a discrepancy of a qualitative physics model, in the case of a selection means-base that selects a particular parameter from a parameter table associated with the model, other parameters may be derived or the model may be considered inconsistent. The particular parameter appears in the equation for the model including the parameter which makes it possible to:-and, in response to the selection means, first assigning means for temporarily assigning a constant value to the parameter, wherein the constant value is possible for the parameter. A propagation means for propagating the constant value through an equation of the model to determine a match, in response to the first assigning means, and a prediction in which a specific constant value of the constant value matches. Until a result or set all possible values for the parameter Repeating means for repeatedly operating the first assigning means and the propagation means until temporarily assigned, and in response to the repeating means, the repeating means coincides when a specific constant value of the constant values is assigned to the parameter. Second assigning means for allocating a specific constant value of the constant values to the parameter when inducing a prediction set, and the selecting means, the first assigning means, the repeating means, and the second allocation for all unknown parameters For a recursive means-model that recursively activates the means and reverts to assigning a new value to a previously assigned parameter whenever the repeating means results in all possible values assigned to the parameter without causing a matching prediction set. Matches values that can be temporarily assigned to unknown parameters in the model, producing a matching set of predictions. If not, the model is deemed inconsistent-mismatch device for determining the qualitative physics model that includes. An apparatus for predicting a value for an unknown parameter associated with a qualitative physics model, comprising: temporary assignment means for temporarily allocating a constant value to the parameter, the constant value being included in a list of possible values for the parameter; In response, propagation means for propagating the constant value through an equation of a model to detect inconsistency, repeating means for repeatedly operating the assignment means and propagation means for all possible values for the parameter, In response, when only the constant value results in a consistent set of predictions for the model, predicting a value for an unknown parameter associated with a qualitative physics model comprising means for permanently assigning a specific value of the constant value to the parameter. Device. An apparatus for determining a discrepancy of a qualitative physics model, comprising: temporary assignment means for temporarily allocating a constant value included in a list of possible values for a parameter to the parameter, and responsive to the temporary assignment means. Propagation means for propagating through equations in the model to detect inconsistencies, repetition means for repeatedly operating the temporary assignment means and propagation means for all possible values for the parameter, and in response to the repetition means, And if no constant value results in a consistent set of predictions for the model, the model comprises means for discarding the model that is considered inconsistent. 18. The apparatus as claimed in claim 17 comprising means for generating one or more qualitative states of a model component and means for providing a user with information indicative of said states. 24. The apparatus of claim 23, further comprising means for comparing the condition with a predetermined set of values assigned for the component. 25. The apparatus of claim 24, wherein said generating means, said providing means, and said comparing means are performed using a computer workstation.