DE10222072A1 - Automatic diagnostic method for dynamic systems involves processing input data in probability network, feeding at least one system characterizing parameter into computation of diagnosis result - Google Patents

Abstract

The method involves determining coefficients characteristic of system output behavior from at least one representative measurement graph, determining discrete system states from the coefficients, assigning the discrete system states as input data to a probability network and processing the input data in the network to compute a diagnosis result. At least one system characterizing parameter is fed into the computation of the diagnosis result. Independent claims are also included for the following: (a) an on-board computer system for a motor vehicle for implementing the inventive method (b) and a diagnostic system for an on-board computer system.

Description

The invention relates to a method and a device for Diagnosis of a dynamic system, especially in the in-vehicle diagnostics (onboard diagnostics) in one Motor vehicle.

In modern motor vehicles today are more functionalities and meeting higher safety and environmental standards required. To meet these requirements, included The vehicles more systems, in turn, more and more complex Subsystems and components included. This applies to mechanics Mechatronics and electronics.

This results in both a greater complexity of the Vehicle, as well as a stronger networking. From the bigger one Complexity follows a larger set of possible causes of error. The stronger networking increases the number of Error impact.

Because many functionalities are not necessary to the It is often also after basic function "driving" to maintain the occurrence of an error possible, the vehicle continues to use, or at least by emergency programs to be able to move to the next workshop.

Monitoring of wear or contamination allows the driver even before the occurrence of a mistake, the prospecting a workshop.

So that these possibilities are correctly evaluated and used can, is a comprehensive picture of the state of possible Causes of faults or wear states during operation of the Vehicle necessary.

Therefore, in the vehicle are small, powerful onboard Diagnostic systems are required, both static and differentiate dynamic symptoms of different sources of error can. For use in systems with dynamic Behavior must include dynamic error symptoms in the evaluation received.

Much of the diagnosis today is in service by the Expert judgment introduced in the troubleshooting. By the Complex systems become this knowledge of experience increasingly larger and therefore no longer completely by one To be human apprehensible. So it is important that at one Diagnosis in a workshop not just the analytical facts but also experience from past errors automatically enter into the diagnostic result.

Even with the onboard diagnosis is such a further Assessment of problems beneficial as the experience many errors in the evaluation of a fault in the vehicle can help. These experiences can take the form of Failure probabilities for a symptom diagnosis in the Vehicle to be transferred.

So that these data are taken into account in a diagnosis it is necessary to work with procedures who can handle blurred knowledge. But with that at the same time an optimal evaluation of sensor information can be done just in systems where a very accurate diagnosis should be made, or at such, at which the evaluation of the dynamics is necessary, one on quantitative models or data based diagnosis meaningful.

An approach to a diagnosis that meets these criteria could be a rating of possible Error candidates with probability networks, eg. B. with Bayesian networks with blurred knowledge about that System can work. This is how a better statement can be made the possible error candidates without further sensor information to reach.

A Bayesian network specifies, for all relations of states of system components (nodes) to each other, the probability of an event H occurring if the event e has occurred:

P (H | e) = x

Through these probabilities x all become causal interdependent states as conditional Probabilities of variables in nodes by one directed graph related to each other. Every node contains different, mutually exclusive states, in sum for each node a probability of 1 must result.

The calculation of Bayesian networks is based on the inversion rule:

With it, statements can be made that are observable Symptoms can be created, and hypotheses H in Set relationship.

Assuming that different statements affect a node that is independent of each other, the relationships of the statements (evidences) e ^k can be summarized as follows:


where α = [P (e ₁ , ..., e _N ] ^{-1 is} a normalization constant.

When a state (for example due to a Observation), the probabilities of the States of all other nodes depending on the conditional Probabilities of these nodes are calculated.

In this calculation, which is called to propagate, are in Recent years have made progress through new algorithms Service. During calculation procedures, which are based on the direct Propagation of the Bayesian network are very complex allows a transformation into from the Bayesian Computable secondary structures, e.g. B. so-called "Junction Trees" a stable and fast calculation. This Secondary structures can be automatically extracted from Bayesian Network generated.

In a Bayesian network, both knowledge can be learned as also be modeled. This applies in both cases for the Structure and conditional probabilities in the network. In order to it is possible to bring all the information into the network which necessary for a diagnosis.

• - Veränderung des Systems
• - Verbesserung der Sensorik
• - Verbesserung des Diagnosealgorithmus

To improve a diagnosis in dynamic systems, there are numerous approaches that can be divided into the following groups:

• - Change of the system
• - Improvement of the sensor technology
• - Improvement of the diagnostic algorithm

Additionally required sensors are expensive and one Changing the system usually brings disadvantages in the Function, the manufacturing and repair effort or at Weight of the systems.

Especially for onboard use with limited memory and Computing capacity is also a quantitative one Diagnostic algorithm is not always applicable as it is generally large computing capacity binds. There are also such models always simplified simulations of reality, and the Data on which they are based, incomplete.

It requires diagnostic approaches that require a maximum of Diagnostic depth with the available sensors and allow limited storage and computing capacity. There much information due to lack of precision of sensors, Models and the system to be diagnosed not sure they are not available in conventional processes considered.

One approach to increasing the depth of diagnosis is therefore the Use of modeling types that are insecure Orient knowledge.

With this additional knowledge can be in question better narrow down the sources of error. It will not be like that renounced early on information. In dependency of Use of the results can be from the Bayesian network assessed probabilities rated differently become. Bayesian networks are used to describe the Relationships between symptoms and their causes in the Used medicine. Especially here a diagnosis is rarely based Base of secure data. So can a greater depth of diagnosis by the processing of information can be achieved, the due to lack of precision only as probabilities available.

US 6,024,705 discloses a method for examining Heart parameters. The measuring curves of the heart parameters are included Help of a Bayesian network classified by To detect cardiac malfunctions. Here are Curve characteristics calculated and with a nearest neighbor Detection algorithm the Bayesian network Assigned to input states.

US 5,133,046 discloses a computer-aided expert System for diagnosing errors in dynamic systems, under Application of a Bayesian network. Here are data values at individual times as input to the Bayesian Network used. The misconduct to be verified independent preprocessing of the data leads either to a inaccurate diagnosis or too big Bayesian networks, which results in a very high computational effort.

However, the method should be in the vehicle, ie in a computer used with low memory and low computing power become. To limit the storage requirements, the calculated output probabilities at each time step be reintroduced into the network, what the computational need further increased.

• - Eine Abhängigkeit des Systemausgangsverhaltens von der Anregung wird nicht berücksichtigt. Auch wenn das Systemverhalten bei der aufgetretenen Anregung unbekannt ist, wird das resultierende Ergebnis verwendet Dies kann zu falschen Diagnoseaussagen führen.
• - Es erfolgt keine Anpassung des Wahrscheinlichkeitsnetzes an bereits vorhandenes Wissen über einen zu diagnostizierenden Systemzustand. Dadurch kann bei einem ungenügenden, z. B. unsicheren Diagnoseergebnis keine verwendbare Aussage über eine Fehlerursache gemacht werden.

Disadvantages of the prior art processes are the following limitations:

• - A dependency of the system output behavior on the excitation is not considered. Even if the system behavior is unknown when the excitation has occurred, the resulting result is used. This can lead to incorrect diagnostic statements.
• - There is no adaptation of the probability network to already existing knowledge about a system state to be diagnosed. This can be at an insufficient, z. B. uncertain diagnosis result no useful statement about a cause of the error can be made.

The invention has for its object a method and a Device for diagnosing dynamic systems to provide the disadvantages of the prior art avoid, and in particular, a system-dependent Enable diagnosis.

According to the invention, the object is achieved by the method and the Device solved according to the independent claims. Special Embodiments are subject of the dependent claims.

The invention is based on the idea that for the creation of a diagnosis result both the input values (suggestions) of a technical system, as well as the initial values (System Output Behavior) Information about the System behavior and thus the system state, in particular Provide error states of the system. An assessment of But the quality of a diagnostic result depends in particular on the suggestion of the system. An output behavior that exactly matches the output behavior, which in a certain system error is expected, can only then Misconduct can be assessed when the suggestion of the system not already identifiable as a potential source of error or if it is known that no suggestions can occur, which can cause such a behavior without the error. So there must be a distinction between a low Error probability of a low diagnostic certainty to be possible. This is possible if the input and the Initial values of the system to be diagnosed as Input data of a probability network, z. B. one Bayesian network initially independent of each other to be viewed as.

In the diagnostic method according to the invention is therefore at least one characterizing the excitation of the system Parameter, in particular a dynamic excitation of the Systems as a characterizing parameter, in the calculation of the Diagnostic result.

According to the invention, the system output behavior characterizing coefficients from at least one of System output behavior representing waveform determined. Out These coefficients are used to determine discrete system states. which as input data a probability network be assigned to. These input data are included in the calculation a diagnosis result with the probability network on.

Also preferred are the excitation of the system characterizing parameter by determining characteristic coefficients of excitation from at least a trace representing the excitation determined, where off the coefficients discrete excitation states are determined which as additional input data of Probability network assigned. The the dynamic Excitation of the system characterizing parameters then go further processed in the calculation as shown of the diagnosis result.

The diagnosis is optimized for a diagnosis Suggested, d. H. it is a time to wait where such a suggestion has occurred and / or is a measure of similarity of excitation of the system to a known one Suggestion, z. B. calculated to an optimized excitation, which in the calculation and / or evaluation of the diagnosis result received.

The similarity of the suggestion then corresponds to one Degree of fulfillment of a quality criterion of the diagnosed System behavior.

The result, which consists of the combination of stimulation and Output behavior is created is a measure of the Likelihood of a misconduct or a Error. If z. Eg the excitation with 1% probability and the system output behavior with 99% probability identified, the result is practically unusable since No information about the system behavior at the occurred excitation exists. Will, however, the suggestion with 99% probability and the initial behavior with 1% Probability identified, so can with 99% security the statement is made that the system is the one to identified initial behavior belonging to misconduct only with 1% probability. This results in the Conversely, that misconduct with 99% he Probability did not occur and thus not further must be considered.

The result of the evaluation of the input data with the Probability network corresponds to the probability that a misconduct has occurred, i. H. a cause of error is present (error hypotheses). When the They can mutually exclude error hypotheses be related to each other by themselves excluding states in a single node be summarized. This reduces z. B. the Probability statement about a fault hypothesis by the Occurrence of a second probable error hypothesis.

The result is a sum of error probabilities that along with a probability value "none of the errors occurred "100% results.

The inventive method is preferred in an onboard Diagnosis in a motor vehicle from an on-board computer system carried out the motor vehicle. The computational effort in the Diagnosis should therefore be as minimal as possible. That's why prefers one of a probability network Bayesian network generated secondary structure used. So that the computational effort in the generation of this Secondary structure does not accumulate in the vehicle, the Transformation of the Bayesian network into one Secondary structure preferred in an external instance outside carried out of the vehicle. The secondary structure can then be used for Calculation along with the original ones Network information to be passed to the vehicle. So is a secure fast and stable calculation possible.

A fault diagnosis in a technical system should recognize as many different errors as possible. At the same time, the computing power and storage needs with cope with the available system resources. To reduce the expense of being operated in the vehicle There is an outsourcing of memory and compute overhead Operations make sense. In an external instance are also available System information in the form of test data, model knowledge and Failure probabilities available. In addition, are Memory and computational resources external to diagnose many Vehicles used, resulting in increased efficiency leads. Therefore, a division of the diagnostic system takes place the vehicle and an external instance (offboard diagnostics). at the external instance can be z. B. an expert center of the motor vehicle manufacturer or one in a workshop act on existing computer.

The test and the creation of a diagnostic result takes place Onboard, new tests for more accurate identification of remaining ones Probable errors can be generated offboard where memory and computing capacities are more favorable and further background knowledge is available.

The probability networks will come with additional data the vehicle sent. The additional data included Information about input and output nodes of the network, Thresholds for the state assignments and parameters for the Preprocessing.

The vehicle stores the diagnostic results, which at the Suggestion with the best similarity value to the desired Were determined and handed over cyclically to the external instance. This is not necessarily a continuous or regular connection between vehicle and external instance needed. An irregular one Data exchange between vehicle and external entity is sufficient.

This results in a diagnostic algorithm that determines its depth of diagnosis gradually in dependence of previous increase probability-based diagnostic statements to progressively connect to one over a longer period of time sufficient or the best possible diagnostic statement zoom elaborate. The required resources remain limited and all kinds of information can be incorporated.

In the illustrated preferred approach is a Adapt the Probability Network to a preliminary one Diagnostic result, with the adjusted Probability network an improved diagnostic result is calculated. Adjusting the Probability Network takes place for. By using new specification data for the Probability network and / or supplement of the Probability network and / or exchange at least of Subnets. The adapted probability network becomes preferably from the external instance to the motor vehicle to hand over. Preference is given to calculating the improved Diagnostic result and adjusting the Probability network to a preliminary diagnosis result as long as repeated until a calculation of a degree of fulfillment a quality criterion of the improved diagnostic result the Fulfilling the quality criterion, with the improved Diagnostic result when not fulfilling the quality criterion as preliminary diagnostic result to which the Probability mesh is adjusted. In which Quality criterion may be z. B. by one Probability of individual error hypotheses that act is exceeded or fallen short of. Very likely Error hypotheses need to be considered further, very much unlikely error hypotheses can against it be ignored.

In calculating the degree of fulfillment of the quality criterion is preferably a characterizing a similarity Parameters of the calculations of the diagnostic results detailed suggestions for at least one optimized suggestion on. This parameter can be z. B. an integral over the difference between the optimized excitation and the currently measured Be a suggestion. The optimized excitation is about a suggestion which is used to determine which one to create Probability network required data was used.

Instead of the network, the complete function and structure of the system to be taken into account in the method according to the invention only the stimulus, the characteristic behavior and the behavior of the faulty systems themselves considered. The information to determine the for the Identification of the behavior necessary probabilities can be modeled as well as based on a quantitative model or based on simulation data be automatically created by learning. This Information will then be in the same network through expert knowledge causally linked. Thus a diagnosis statement on a wider database is available without the number of input nodes enlarge, a simple preprocessing takes place. By the combination, z. B. an integration of diagnosing measurement data by means of a Trapezium integration method with a probability network the computational effort and memory requirements are reduced.

While the number of nodes of a Bayesian network the Size of a mesh is only linearly influenced, the number goes up Inputs into a node exponentially in the network size. Therefore, the network is preferably designed so that between the Input node and the node, which is a diagnostic result represent intermediate nodes exist. They represent the Similarity with which the identified waveforms the correspond to learned waveforms. If several signals which represent excitation and / or the output behavior is to identify the entire excitation process and Output behavior for a fault, a common node assigned. The simplest creation of a Signal identification can be achieved by teaching the signal over take place over a certain period of time. The shorter the Training time and the lower the dynamics of the signals, the more the result is more exact. The longer the learning time and ever greater the dynamics, the greater the tolerance in possible Deviations.

Various models or simulation data can be taught-in become. It is important that the formed integrals at the different error cases and against the Distinguish nominal behavior as much as possible. That's why the selection of data by an examination of the Diagnosability during preprocessing.

The quality of a diagnostic statement will also depend on the type of Discretization of the integration values of the measured data as States of the input nodes influenced. Depending on whether the Value range of the measured data around a mean value of the Value range of the learning data for discretization divided will, or depending on the maximum and minimum Learning data is split, the result is absolute Probability, or only relative to the others Evaluate errors.

A preferred way of discretization involves Dividing the value range of the learning data into uniform Discretization states, and adding two states, which include all measurement data, the larger or the smaller are.

If the learning data with offset errors by model or Sensor inaccuracies are affected, this method is not sufficient to give an absolute probability of error obtained there with low dynamics of the integral value of the Range of values that represents an error becomes very narrow, while he gets bigger at high dynamics and thus one Offset error is no longer significant.

So only one probability of error is relative to the others Error hypotheses possible.

An absolute statement about how exactly a signal a known behavior is possible only if the states of the signal detection as a function of expected Uncertainties are defined precisely.

The result of a diagnosis depends on the value of the data. This corresponds to the probability that the Input data match the learning data. They go as Probability of a probability or as one Second order probability in the network. With this Value of the data is a further processing or a Combination with other statements, eg. B. Expert knowledge possible.

• - Verfahren mit nicht parametrischen Modellen basieren auf dem "black box" Prinzip. Das bedeutet, sie bilden nur das Verhalten eines Systems nach, nicht aber dessen Struktur oder dessen innere Zustände. Nicht parametrische Modelle können deshalb leicht eingelernt, aber nur aufwändig aus Systemwissen oder Expertenwissen modelliert werden. Sie bestehen aus neuronalen Netzwerken oder Formen von Fuzzy Logik.
• - Logik-basierte Diagnoseverfahren arbeiten mit der Systemstruktur und dem qualitativen Komponentenverhalten. Mit ihnen können Zustandsautomaten oder logische Sätze zur Überprüfung von Hypothesen generiert werden.
• - Analytische Verfahren arbeiten auf Basis der Struktur und des mathematisch beschriebenen Komponentenverhaltens. Ein Diagnosesystem, das mit einem solchen Verfahren arbeitet, ist in der Diagnosetiefe nur beschränkt durch die Sensoren und durch Fehler in der Systembeschreibung. Da jede Modellbeschreibung komplexer Systeme Vereinfachungen enthalten muss, sind die Ergebnisse abhängig von der Modellgenauigkeit. Ausreichend genaue Modelle, die Soll- und Fehlverhalten nachbilden sind aber beim Einsatz im Fahrzeug oft zu komplex, um berechnet zu werden.

The preliminary diagnosis result can be z. B. by the following methods according to the prior art or even by a first approach of the method according to the invention are determined with a first probability network:

• - Procedures with non-parametric models are based on the "black box" principle. This means that they only model the behavior of a system, not its structure or its internal states. Non-parametric models can therefore be easily learned in, but only elaborately modeled from system knowledge or expert knowledge. They consist of neural networks or forms of fuzzy logic.
• - Logic-based diagnostics work with system structure and qualitative component behavior. They can be used to generate state machines or logical sentences to test hypotheses.
• Analytical methods work on the basis of the structure and the mathematically described component behavior. A diagnostic system using such a method is limited in the depth of the diagnosis only by the sensors and errors in the system description. Since every model description of complex systems must include simplifications, the results depend on the model accuracy. However, sufficiently accurate models that mimic setpoint and misconduct are often too complex to use in the vehicle to calculate.

• - Es ergibt sich ein Diagnosealgorithmus, der seine Diagnosetiefe schrittweise in Abhängigkeit vorhergehender wahrscheinlichkeitsbasierter Diagnoseaussagen vergrößern kann. Dadurch ist eine optimale Diagnoseaussage über einen längeren Zeitraum iterativ berechenbar.
• - Die benötigten Ressourcen sind begrenzt und beliebige Arten von Informationen können in die Diagnose einfließen. Das Diagnoseergebnis kann Teil eines größeren Wahrscheinlichkeitsnetzes werden welches z. B. aus dem Wissen über die Fehler Aktionen ableitet.
• - Ein Ähnlichkeitsmaß der Systemanregung kann als Qualitätskriterium des Ergebnisses gewertet werden. Es müssen nicht alle Symptome gleich behandelt werden. Es wird nicht von einer immer identischen Anregung ausgegangen.
• - Ein schrittweises Verringern der zu berücksichtigenden Fehlerhypothesen und eine gleichzeitige Optimierung der neuen Diagnose bezüglich der verbleibenden Resthypothesen kann erfolgen. Wurde ein Anregungsverlauf nicht identifiziert, wird eine darauf folgende Diagnose auf einen anderen Anregungsverlauf ausgelegt. Fällt eine Hypothese weg, ist diese nicht mehr Kriterium für die Wahl einer neuen Anregung.
• - Unterschiedlichen Fehlerhypothesen können unterschiedliche Netzwerkfragmente zugewiesen werden, was bei einer fehlerspezifischen Schwellwertdefinition notwendig ist.

By the method according to the invention and the devices for carrying out the method, the disadvantages of the prior art are avoided. In particular, the following advantages are realized:

• A diagnostic algorithm results which can increase its depth of diagnosis step by step as a function of previous probability-based diagnostic statements. As a result, an optimal diagnosis statement can be calculated iteratively over a longer period of time.
• - The resources required are limited and any type of information can be included in the diagnosis. The diagnostic result can be part of a larger probability network which z. B. derives from the knowledge about the errors actions.
• - A similarity measure of the system excitation can be considered as the quality criterion of the result. Not all symptoms need to be treated equally. It is not assumed that there is always an identical suggestion.
• - A gradual reduction of the error hypotheses to be considered and a simultaneous optimization of the new diagnosis with respect to the remaining residual hypotheses can take place. If an excitation process has not been identified, a subsequent diagnosis is designed for a different excitation process. If one hypothesis disappears, it is no longer a criterion for choosing a new stimulus.
• - Different error hypotheses can be assigned different network fragments, which is necessary for an error-specific threshold definition.

The present invention and preferred embodiments of the Invention will be described below with reference to figures.

Fig. 1 shows the basic structure of a diagnostic system according to the invention.

FIG. 2 shows the method according to the invention with regard to the consideration of values which characterize an excitation of the system to be diagnosed in calculating the diagnosis result.

Fig. 3, the inventive method, especially the adaptation of the probability network to a preliminary diagnosis result in interaction with an external entity FIG.

FIG. 1 shows a diagnostic system according to the invention and the sequence of the diagnostic method according to the invention. The method according to the invention is applied to a dynamic technical system 100 . An excitation, which is represented by the measuring curve 110 , acts on the system. The system responds to the excitation with a system output behavior represented by a trace 120 . The measured curves are each drawn in a coordinate system. The diagnostic system according to the invention consists of an on-board computer 200 and an external instance 300 . The on-board computer has determining means 250 of characteristic parameters of the. Excitation and determination means 260 of characteristic coefficients of the system output behavior. These determination means are set up from the characteristic parameters or the coefficients of the system output behavior to determine discrete system states 60 , or to determine discrete excitation states 50 . This is done by dividing the size plotted on the abscissas of the coordinate systems into integration areas. Preferably, this size is the time. The excitation is divided into four integration areas 1 , 2 , 3 , 4 . By integration in these areas, the excitation characterizing parameters are determined. The determined from these parameters excited state is z. B. as indicated in the figure, a discrete state "bcca". Also, the system output behavior representing waveform is divided into integration areas 5 , 6 , 7 , 8 . Discrete system states are determined 60 from integrals calculated in these areas which form coefficients characterizing the system output behavior. In the figure, z. B. a discrete state "ddba" determined. The on-board computer 200 further has an evaluation unit 210 and means 22 for assigning the discrete system states as input data of a probability network 10 . The probability network 10 is represented symbolically in the figure as network hatching. Furthermore, means 21 for assigning the discrete excitation states are provided as input data of the probability network. The input data are processed by the evaluation unit in the probability network 70 , 80 . In this case, a diagnostic result 30 is calculated. In the example of the figure, three error hypotheses H1, H2 and H3 are tested. As a result of the diagnosis z. Example, that the hypothesis H2 corresponds to a probability of 90% for the given input data, the hypothesis H1 has a probability of 10% and the hypothesis H3 with 0% probability, impossible to apply in the given input data. The system according to the invention furthermore has means for calculating a degree of fulfillment of a quality criterion of the diagnostic result 90 . Such a quality criterion can, for. B. the degree of security of a tested hypothesis. Such a quality criterion serving as a security level can, for. For example, the requirement that the known excitation process was met with 95% probability and / or the diagnostic result of an error is above 90% or below 10%. If the desired degree of fulfillment is not achieved, it is necessary to repeat the diagnosis under improved conditions. This can be done by modifying the probability network and calculating a new diagnostic result, processing the same and / or changed, e.g. B. done refined input data in the modified probability network. For this purpose, a probability network 11 adapted to the preliminary diagnostic result, which has not met the quality criterion, is generated in an external instance 300 . The adapted probability network is symbolized in the figure as a checkered surface. The modified probability network is transferred 40 to the on-board computer. The illustrated diagnostic sequence is repeated until a diagnostic result has been calculated, which has the desired degree of fulfillment of the quality criterion.

In FIG. 2, the method according to the invention is described with regard to the consideration of parameters which characterize excitation u of the system to be diagnosed in calculating the diagnosis result. In the example shown, two system outputs with the associated output variables y1 and y2 are taken into account in the diagnosis. The excitation and the resulting system output behavior are shown as measurement curves (signal curve) as a function of time t.

Through preprocessing, different parts of a Waveform processed so that they each as a discrete States the input data of a probability network can form. The system output behavior and / or the Excitation characterizing coefficients are preferred by dividing the system output behavior and / or the excitation - representing measuring curves in Partial integration areas and determine integral values of Partial integration areas determined.

In the time domain, this can be done by computing integrals, e.g. B.


take place, wherein the time is divided into several integration areas between times ti or averaging occurs. In the figure, the integrals A, B, C, and D are determined for the excitation u. From the history of the system output characteristics, the integrals E, F, G, and H are calculated for the output variable y1 and I, J, K, and L for the output variable y2. In the frequency domain, this can be done by the deviation of the integral or mean over the energy density in certain areas.

By supplementing the preprocessing with a Transformation in the frequency domain, z. B. by a fast Fourier transformation, the time information from the signal be filtered out. This will be the System output behavior and / or the excitation performing Traces before determining the system output behavior and / or the excitation characterizing coefficients in the Frequency space transformed.

The amplitudes of the energy bands in the frequency domain become by a slight shift of the considered time window of the signal is less affected than the amplitudes of the signal Time signal itself. This will determine the the Measuring curves characterizing coefficients of the concrete Limitations of the integration areas influenced less. The Number of integration areas can therefore be in the frequency domain, with sufficiently accurate characterization of the trace, be less than in the time domain. There is no danger that an occurring signal is not displayed.

Depending on the value range of the specific coefficients, discrete system states AD are then determined. The determination of the discrete system states and / or excited states is performed by assigning the integral values to the discrete system states and / or excited states. For example, for the division into three states a, b and c of the integral value A, the limit 1 and the limit 2 are defined within the value range of A:

Depending on the number of integration areas, one sets Trace representing discrete state then from a corresponding number of substates AD together. The States become the probability network as input data assigned. In the figure, the states become AD, BD, CD and DD the system excitation, the states Ed to Hd of the system output y1 and Id to Ld of the system output y2 the input node Ni assigned to the probability network. The nodes N5 to N12 are the input nodes for the error hypothesis FH1 and the nodes N13 to N20 the input nodes for the Error hypothesis FH2.

If, as in the example shown several inputs and / or Outputs y1, y2 are present, the probabilities become the corresponding output behavior of the system corresponding suggestions of the system to an error input or output behavior summarized. This is in the figure represented by the nodes Nfv1 and Nfv2.

By combining the probability of the occurrence of various errors in a particular excitation and the different system output behavior then the probability that various causes of errors have occurred (error hypotheses) in the same probability network can be determined. In the figure this is shown for two error hypotheses (FH1 and FH2). The discretized system output behaviors y1 and y2 are combined to the two output behaviors Fv1 and Fv2. Assuming that only one error has occurred, then in the network the discretized output behaviors are assigned as input data of the probability network. Through network propagation, the probabilities F1, F2 are then calculated so that the error hypothesis 1 or the error hypothesis 2 applies to an excitation u.

The value (quality) W of the calculated data of this Diagnosis result can be identified by the identification of the define the learned input signal. The quality of the Diagnosis result depends on the similarity of the Excitation, which led to the system output behavior y1, y2 has to do in determining the conditional probabilities used suggestions. This similarity can be in the Calculating a degree of fulfillment of a quality criterion of Result of the diagnosis.

To limit the memory and computing capacity is only one minimal network in the vehicle. For more detailed Checking can be new networks or network fragments in the Vehicle depending on the calculated diagnostic results be reloaded.

FIG. 3 describes the method according to the invention, in particular the adaptation of the probability network to a preliminary diagnosis result. The diagnosis and the generation of probability networks is performed in successive, generally different time intervals Δt _i . In the example shown, the probability network (probabilistic model) is adapted twice to a provisional diagnostic result. The network propagation is carried out in the time intervals Δt _i , Δt ₃ and Δt ₅ , while the measurement of the excitations n and the system output behavior y in the time intervals .DELTA.t ₂ and .DELTA.t ₄ and in a time interval before t _{0 is} made.

• - Zunächst kann eine vorläufige Diagnose durchgeführt werden. Diese kann durch einfache Algorithmen, z. B. Schwellwertüberwachung oder eine erste Diagnose mit einem Startwahrscheinlichkeitsnetz erfolgen, welches überprüft, ob ein für das nicht Auftreten eines Fehlers charakteristisches Systemausgangsverhalten nach einer charakteristischen Anregung auftritt.
• - Ist ein Fehler aufgetreten findet entweder eine Grobdiagnose statt, bei der eine große Anzahl von Fehlern mit kleinen Unternetzwerken überprüft wird oder eine Diagnose, welche nur einen Teil der möglichen Fehler überprüft. Die Grobdiagnose kann auch verschiedene Fehlerfälle mit ähnlichen Fehlersymptomen als Gruppe durch je ein Unternetzwerk gleichzeitig prüfen. Eine derartige Grobdiagnose ist in der Figur mit dem Wahrscheinlichkeitsmodell in der Auswerteeinheit, welche symbolisch im Zeitbereich Δt₁ dargestellt ist angedeutet. Es werden zunächst vier Fehlerhypothesen getestet. Als Grobdiagnoseergebnis ergeben sich für die den Hypothesen entsprechenden Fehler Wahrscheinlichkeiten von 29%, 40%, und zwei mal 2%. Das Diagnoseergebnis wird von der externen Expertenzentrale ausgelesen.
• - Das weitere Vorgehen hängt nun bei jedem Schritt davon ab, ob eine ähnliche Anregung im vorhergehenden Schritt aufgetreten ist und ob das Diagnoseergebnis insgesamt ausreicht. Es kann in diesem Schritt sowohl das gesamte Netz und/oder die gesamten Spezifikationsdaten ausgetauscht werden, als auch nur Teile davon. Im dargestellten Beispiel wird eine Neukonfiguration des Wahrscheinlichkeitsnetzes durch Austausch eines Submodells durchgeführt. Das neue Submodell wird von der externen Expertenzentrale an die Auswerteeinheit übergeben. Danach erfolgt ein weiterer Diagnoseschritt, dargestellt im Zeitbereich Δt₃. In diesem Schritt ist das Wahrscheinlichkeitsnetz auf den Test von drei Fehlerhypothesen reduziert. Das Diagnoseergebnis verbessert sich auf eine Sicherheit von 70% für einen bestimmten Fehler. Das Diagnoseergebnis wird wiederum von der externen Expertenzentrale aus der Auswerteeinheit ausgelesen.
• - Falls eine Anregung, welche im aktuellen Netz bekannt ist nach einer vorgegebenen Zeit nicht auftritt und damit in nächster Zeit als unwahrscheinlich angenommen werden kann, muss ein neues Netz mit zugehöriger Spezifikation auf eine andere Anregung ausgelegt werden. In diesem Falle wird aus den verbleibenden Daten, d. h. den verbleibenden optimierten Anregungen, für welche Netzwerke zur Verfügung stehen, der Verlauf ausgewählt, der die beste Unterscheidung zwischen den zu prüfenden Fehlerhypothesen erwarten lässt. Als Kriterium kann z. B. die Summe der Differenzen zwischen den Integralwerten der zu Fehlerhypothesen passenden Systemausgangsverhalten dienen. Die Fehlerdiagnose erfolgt mit dem derartig ausgewählten Netz im Fahrzeug erneut.
• - Ist die Anregung mit ausreichender Ähnlichkeit zu bekannten Anregungen aufgetreten, das Diagnoseergebnis jedoch nicht ausreichend, hinsichtlich dessen Erfüllungsgrades eines Qualitätskriteriums, so sind die Diagnoseergebnisse des vorhergehenden Schrittes nicht zufriedenstellend. Dann werden Fehlerhypothesen die oberhalb eines Schwellwertes liegen in einem neuen Netzwerk in einem weiteren Schritt erneut betrachtet. Nicht zufriedenstellend ist ein Diagnoseergebnis dann, wenn entweder zu viele Einzelfehler wahrscheinlich sind oder kein Einzelfehler wahrscheinlich genug ist, um ihn als Diagnoseergebnis auszugeben zu können.
• - Falls nur ein Teil der Fehlerhypothesen im vorhergehenden Schritt überprüft wurde, wird nun ein Netz für noch nicht überprüfte Fehlerhypothesen aufgestellt und an das Fahrzeug übergeben.
• - Falls Gruppen von Fehlerhypothesen in einem Netzwerk eingelernt wurden sind die Fehlerhypothesen der verbleibenden Gruppen oder Untergruppen nun durch einzelne Unternetzwerke zu überprüfen.
• - Die Diagnose wird beendet wenn die Anregung mit ausreichender Ähnlichkeit aufgetreten ist und das Diagnoseergebnis ausreichend ist oder kein besseres Ergebnis mehr erwartet werden kann. Das Diagnoseergebnis kann dann im Fahrzeug angezeigt werden und/oder an, z. B. eine Werkstatt übermittelt werden. In der Figur ist dies nach dem Berechnen des Diagnoseergebnisses mit dem Wahrscheinlichkeitsnetz im Zeitbereich Δt₅ der Fall. Dieses Wahrscheinlichkeitsnetz wurde gegenüber dem Wahrscheinlichkeitsnetz, welches der Diagnose im Zeitbereich Δt₃ zugrunde lag, durch Austausch des Wahrscheinlichkeitsnetzes an das vorhergehende Diagnoseergebnis angepasst. Es resultieren die Fehlerwahrscheinlichkeiten 10% und 90%

Bezugszeichenliste 1, 2, 3, 4 Integrationsbereiche der die Anregung darstellenden Messkurve zum Bestimmen von die Anregung charakterisierenden Koeffizienten
5, 6, 7, 8 Integrationsbereiche der das Systemausgangsverhalten darstellenden Messkurve zum Bestimmen von das Systemausgangsverhalten charakterisierenden Koeffizienten
10 Wahrscheinlichkeitsnetz
11 an ein vorläufiges Diagnoseergebnis angepasstes Wahrscheinlichkeitsnetz
21 Mittel zum Zuweisen diskreter Anregungszustände als Eingangsdaten des Wahrscheinlichkeitsnetzes
22 Mittel zum Zuweisen diskreter Systemzustände als Eingangsdaten des Wahrscheinlichkeitsnetzes
30 Diagnoseergebnis
40 Übergabe des an ein vorläufiges Diagnoseergebnis angepassten Wahrscheinlichkeitsnetzes
50 Bestimmen diskreter Anregungszustände
60 Bestimmen diskreter Systemzustände
70, 80 Verarbeiten der Eingangsdaten
90 Berechnen eines Erfüllungsgrades eines Qualitätskriteriums
100 dynamisches technisches System
110 Anregung darstellenden Messkurve
120 Systemausgangsverhalten darstellenden Messkurve
200 Bordcomputersystem
210 Auswerteeinheit
250 Bestimmungsmittel von charakteristischen Parametern der Anregung
260 Bestimmungsmittel von charakteristischen Koeffizienten des Systemausgangsverhaltens
300 externe Instanz
In general, the diagnostic method according to the invention can be divided into the following method steps:

• - First, a preliminary diagnosis can be made. This can be achieved by simple algorithms, eg. B. Threshold monitoring or a first diagnosis with a start likelihood network, which checks whether a non-occurrence of a fault characteristic system output behavior occurs after a characteristic excitation.
• - If an error has occurred, either a rough diagnosis is carried out, in which a large number of errors are checked with small subnetworks or a diagnosis, which only checks a part of the possible errors. The rough diagnosis can also simultaneously check different error cases with similar error symptoms as a group by one subnetwork at a time. Such a rough diagnosis is indicated in the figure with the probability model in the evaluation unit, which is shown symbolically in the time domain Δt ₁ . First, four error hypotheses are tested. As a rough diagnosis result for the hypothesis corresponding errors probabilities of 29%, 40%, and two times 2%. The diagnostic result is read out by the external expert center.
• - The next procedure now depends on each step on whether a similar excitation has occurred in the previous step and whether the total diagnostic result is sufficient. In this step, both the entire network and / or the entire specification data can be exchanged, as well as only parts thereof. In the example shown, a reconfiguration of the probability network is performed by exchanging a submodel. The new submodel is transferred from the external expert center to the evaluation unit. This is followed by another diagnostic step, represented in the time domain Δt ₃ . In this step, the probability network is reduced to the test of three error hypotheses. The diagnostic result improves to a 70% certainty for a given error. The diagnostic result is again read out of the evaluation unit by the external expert center.
• - If an excitation, which is known in the current network does not occur after a predetermined time and thus can be considered unlikely in the near future, a new network with associated specification must be interpreted to another suggestion. In this case, from the remaining data, ie the remaining optimized excitations for which networks are available, the course is selected that gives the best distinction between the error hypotheses to be tested. As a criterion z. For example, the sum of the differences between the integral values of the system output characteristics suitable for error hypotheses. The fault diagnosis is carried out again with the network selected in the vehicle.
• - If the excitation occurred with sufficient similarity to known suggestions, but the diagnostic result is not sufficient in terms of its degree of fulfillment of a quality criterion, the diagnostic results of the previous step are not satisfactory. Then error hypotheses that are above a threshold value are considered again in a new network in a further step. A diagnostic result is unsatisfactory if either too many individual errors are probable or if no single error is likely enough to be able to output it as a diagnostic result.
• - If only part of the fault hypotheses were checked in the previous step, a network for fault hypotheses that have not yet been checked is now set up and transferred to the vehicle.
• - If groups of error hypotheses have been learned in a network, the error hypotheses of the remaining groups or subgroups must now be checked by individual subnetworks.
• - The diagnosis is terminated when the stimulation with sufficient similarity has occurred and the diagnostic result is sufficient or no better result can be expected. The diagnostic result can then be displayed in the vehicle and / or on, for. B. a workshop to be transmitted. In the figure, this is the case after the calculation of the diagnosis result with the probability network in the time domain Δt ₅ . This probability network was compared with the probability network, which was the basis of the diagnosis in the time domain .DELTA.t ₃ , adapted by replacing the probability network to the previous diagnosis result. This results in the error probabilities 10% and 90%

LIST OF REFERENCE NUMERALS 1 , 2 , 3 , 4 Integration regions of the excitation representing trace for determining excitation characterizing coefficient
5 , 6 , 7 , 8 are integration ranges of the system output characteristic waveform for determining coefficients characterizing the system output behavior
10 probability network
11 probability network adapted to a preliminary diagnosis result
21 means for assigning discrete excitation states as input data of the probability network
22 means for assigning discrete system states as input data of the probability network
30 Diagnostic result
40 Transfer of the probability network adapted to a preliminary diagnostic result
50 determining discrete excitation states
60 determining discrete system states
70 , 80 Processing the input data
90 Calculating a degree of fulfillment of a quality criterion
100 dynamic technical system
110 excitation-representing trace
120 system output behavior representative trace
200 on-board computer system
210 evaluation unit
250 Determination means of characteristic parameters of excitation
260 Determining means of characteristic coefficients of the system output behavior
300 external instance

Claims (17)

1. Automatic diagnostic method for a dynamic technical system, which responds to an excitation with a dynamic system output behavior, with the method steps: Determining coefficients characterizing the system output behavior from at least one measurement curve representing the system output behavior, Determining discrete system states from the coefficients, Assign the discrete system states as input data of a probability network and Processing the input data in the probability network to calculate a diagnostic result; characterized in that at least one parameter characterizing the excitation of the system is included in the calculation of the diagnostic result. 2. A diagnostic method according to claim 1, characterized in that the excitation of the system characterizing parameters by determining characteristic coefficients of excitation at least one trace representing the excitation be determined, with the excitation to a dynamic stimulus acts and from the coefficients discrete excitation states are determined which as Assigned input data of the probability network become. 3. The diagnostic method according to at least one of claims 1 to 2, characterized in that the System output behavior and / or excitation characterizing coefficients by dividing the the system output behavior and / or the excitation representing measuring curves in partial integration areas and determine integral values of the subintegration areas and determining discrete system states and / or excitation states by assigning the integral values to the discrete system states and / or excited states he follows. 4. The diagnostic method according to at least one of claims 1 to 3, characterized in that the diagnosis at least performed a stimulation optimized for the diagnosis or that a similarity measure of the excitation of the system to an optimized stimulus in the calculation of the Diagnostic result received. 5. The diagnostic method according to at least one of claims 1 to 4, characterized in that it is in the Probability network around a Bayesian network or around one generated by a Bayesian network Secondary structure acts. 6. The diagnostic method according to at least one of claims 1 to 5, characterized in that the System output behavior and / or the excitation performing Traces before determining the System output behavior and / or excitation characterizing coefficients in the frequency domain be transformed. 7. The diagnostic method according to at least one of claims 1 to 6 characterized in that an adaptation of the Probability network to a preliminary Diagnostic result is performed, with the adjusted Probability network an improved diagnostic result is calculated. 8. A diagnostic method according to claim 7, characterized in that the provisional Diagnostic result by analyzing the system output behavior with a logic based diagnostic method and / or a analytical procedures and / or with a Probability network is determined. 9. A diagnostic method according to at least one of claims 7 to 8th, characterized in that the adaptation of the Probability network by using new ones Specification data for the probability network and / or supplementing the probability network and / or Exchange of at least subnetworks takes place. 10. The diagnostic method according to at least one of claims 7 to 9 characterized in that the computing of the improved Diagnostic result and adjusting the Probability network to a preliminary Diagnostic result is repeated until a calculation a degree of fulfillment of a quality criterion of the improved diagnostic result the fulfillment of the Quality criterion results, with the improved Diagnostic result if the quality criterion is not fulfilled as a preliminary diagnostic result to which the Probability mesh is adjusted. 11. A diagnostic method according to claim 10, characterized in that in calculating the Degree of fulfillment of the quality criterion a Similarity of characterizing parameters of the Calculations of the diagnostic results of incoming suggestions at least one optimized suggestion is received. 12. A diagnostic method according to at least one of claims 1 to 11 characterized in that the method at a Onboard diagnosis in a motor vehicle of one On-board computer system of the motor vehicle is performed. 13. A diagnostic method according to claim 12 and at least one of Claims 7 to 11, characterized in that the adapted Probability network from an external instance to the Motor vehicle is handed over. 14. On-board computer system of a motor vehicle, set up to carry out the diagnostic method according to at least one of claims 1 to 13, on a dynamic technical system, which responds to an excitation with a dynamic system output behavior, with Means for determining characteristic coefficients of the system output behavior, Means for determining discrete system states from the coefficients, - Means for assigning the discrete system states as input data of a probability network of an evaluation unit for processing the input data in the probability network for calculating a diagnosis result characterized in that the evaluation unit for processing the input data in the probability network is set up to calculate a diagnosis result dependent on at least one parameter or condition characterizing the excitation of the system. 15. on-board computer system according to claim 14, characterized in that means for modifying the Probability network in a to a provisional Diagnostic result adapted probability network are provided. 16. on-board computer system according to at least one of claims 14 to 15, characterized in that means for calculating a Degree of fulfillment of a quality criterion of the Diagnosis result are provided. 17. Diagnostic system with an on-board computer system after at least one of claims 15 to 16, characterized in that an external instance exists is where the external instance is set to on preliminary diagnosis result adjusted Probability network to generate and to the Hand over on-board computer system and the on-board computer system set up is the adapted probability network too take over.