DE102020107528A1 - System for fault diagnosis and / or fault prognosis in motor vehicles - Google Patents

Abstract

The diagnostic system according to the invention for anomaly detection and / or error prognosis in motor vehicles, which classifies both known errors and unknown errors on the basis of behavior vectors, is equipped with an internal vehicle diagnostic unit (on-board unit) and / or with a Vehicle-external diagnostic unit (off-board unit) equipped. According to the invention, at least one of the two diagnostic units contains a first database in which behavior vectors known for known operating modes are stored. Furthermore, at least one of the two diagnostic units contains a diagnostic control unit with an evaluation algorithm for anomaly detection and / or for error prognosis, which has a vector similarity test for comparing determined behavior vectors with known behavior vectors in the database in order to identify unknown behavior vectors identify.

Description

The invention relates to a system for error diagnosis and / or error prognosis in motor vehicles with an evaluation module which classifies errors on the basis of behavior vectors.

Technical background:

In order to improve the performance of diagnosis and prognosis ( PHM ) There are a number of metrics to evaluate methods. The methods based on behavior vectors outperform existing methods in terms of variant handling and the ability to analyze highly complex systems. In addition, the concept of behavior vectors is not only limited to vehicle diagnosis, but can also address issues of prognosis ( PHM ) as well as the quality analysis are applied. It is to be expected that the development costs will decrease due to the high degree of automation and reusability of the models compared to existing methods. Other outstanding features of the methods are the ability to detect unknown errors and cross-domain applicability.

The known methods on which the classification is based are provided in the table in 9 .Error! Reference source not found.

It is the object of the invention to develop such known methods with regard to the detection of unknown errors.

This object is achieved by the features of claim 1. Advantageous further developments of the invention are the subjects of the dependent claims.

The inventive diagnostic system for anomaly detection and / or error prognosis in motor vehicles, which classifies both known errors and unknown errors on the basis of behavior vectors, is equipped with an internal vehicle diagnostic unit (on-board unit) and / or with a Vehicle-external diagnostic unit (off-board unit) equipped. According to the invention, at least one of the two diagnostic units contains a first database in which behavior vectors known for known operating modes are stored. Furthermore, at least one of the two diagnostic units contains a diagnostic control unit with an evaluation algorithm for anomaly detection and / or for error prognosis, which has a vector similarity test for comparing determined behavior vectors with known behavior vectors in the database in order to identify unknown behavior vectors identify.

Identified unknown behavior vectors are preferably stored in a second database in order to store them in the first database by means of an evaluating expert system either to detect errors in a previously known operating mode or as new known behavior vectors in a then new known operating mode.

In an advantageous development of the invention, the diagnostic control unit is also designed in such a way that a failure probability is predictable after a tendency of a behavior vector to deviate from an operating mode has been detected and quantified over time with regard to angle, position and / or length.

• 1 eine detaillierte schematische Darstellung des erfindungsgemäßen Gesamtkonzepts,
• 2 die qualitative Einordnung der Erfindung zur Klassifikation gemäß dem Stand der Technik in Bezug auf handhabbare Systemkomplexität abhängig von den Kosten und dem Aufwand zur Methodenerstellung,
• 3 die qualitative Einordnung der Erfindung zur Klassifikation gemäß dem Stand der Technik in Bezug auf handhabbare Variantenvielfalt abhängig von den Kosten und dem Aufwand zur Methodenerstellung,
• 4 einen Überblick der Systemtopologie eines erfindungsgemäßen Diagnosesystems,
• 5 einen zwei-dimensionalen (f₁, f₂) Zustandsraum mit zwei aufeinander folgenden Zuständen und dem korrespondierenden Verhaltensvektor,
• 6 einen zwei-dimensionalen Zustandsraum mit bekannten ähnlichen Verhaltensvektoren (u₁, u₂, u₃) und beobachteten Verhaltensvektoren $\left({\text{u}}_2^{\ast },{\text{u}}_4^{\ast }\right)$
  
  für die Fehler- und Anomalie Detektion,
• 7 Verhaltensvektoren für den Fehlerprognose und Health Management Use-Case (=Anwendungsfall) und
• 8 Verhaltensvektoren für den Anwendungsfall für die Qualitätssicherung („QA Use-Case“).

The invention is based on the following considerations, which are also explained in more detail with reference to the drawing. It shows

• 1 a detailed schematic representation of the overall concept according to the invention,
• 2 the qualitative classification of the invention for classification according to the state of the art in terms of manageable system complexity depending on the costs and the effort involved in creating the method,
• 3 the qualitative classification of the invention for classification according to the state of the art with regard to manageable variety of variants depending on the costs and the effort involved in creating the method,
• 4th an overview of the system topology of a diagnostic system according to the invention,
• 5 a two-dimensional (f ₁ , f ₂ ) state space with two successive states and the corresponding behavior vector,
• 6th a two-dimensional state space with known similar behavior vectors (u ₁ , u ₂ , u ₃ ) and observed behavior vectors $\left({\text{<figure-callout id="2" label="u" filenames="DE102020107528A1\_0005.png" state="{{state}}">u</figure-callout>}}_2^{\ast },{\text{u}}_{\mathrm{4th}}^{\ast }\right)$
  
  for error and anomaly detection,
• 7th Behavioral vectors for error prognosis and health management use case (= use case) and
• 8th Behavioral vectors for the use case for quality assurance (“QA use case”).

• Das Diagnosesystem zur Anomalie-Erkennung und/oder Fehler-Prognose in Kraftfahrzeugen, das auf Basis von Verhaltensvektoren u sowohl bekannte Fehler als auch unbekannte Fehler klassifiziert, weist eine Kraftfahrzeug-internen Diagnose-Einheit 1 (On-Board Einheit) und hier optional eine Kraftfahrzeug-externe Diagnose-Einheit 2 (Off-Board) auf. Bei diesem Ausführungsbeispiel ist in der ersten Diagnose-Einheit 1 eine erste Datenbank 3 enthalten, in der zu bekannten Betriebsmodi m_j bekannte Verhaltensvektoren u(m_j ) abgespeichert sind. Weiterhin ist hier in der ersten Diagnose-Einheit 1 eine Diagnose-Steuereinheit 4 (Diagnose-ECU) mit einem Auswertealgorithmus zur Anomalie-Erkennung und/oder zur Fehler-Prognose enthalten, der einen Vektoren-Ähnlichkeitstest zum Vergleich ermittelter Verhaltensvektoren u mit bekannten Verhaltensvektoren u(m_j ) der Datenbank 3 aufweist, um unbekannte Verhaltensvektoren u(m_j )* zu identifizieren.

In 1 an exemplary embodiment for the following basic idea according to the invention is shown:

• The diagnostic system for anomaly detection and / or error prognosis in motor vehicles, which is based on behavior vectors u Both known faults and unknown faults classified, has a motor vehicle-internal diagnostic unit 1 (on-board unit) and here optionally a motor vehicle-external diagnostic unit 2 (off-board). In this exemplary embodiment, the first diagnosis unit 1 contains a first database 3 in which operating modes are known m _j known behavior vectors u ( m _j ) are saved. The first diagnostic unit 1 also contains a diagnostic control unit 4 (diagnostic ECU) with an evaluation algorithm for anomaly detection and / or for error prognosis, which performs a vector similarity test for comparing determined behavior vectors u with known behavior vectors u ( m _j ) of the database 3 to identify unknown behavior vectors u ( m _j )* to identify.

The identified unknown behavior vectors u ( m _j ) * are used in the exemplary embodiment according to 1 forwarded to the second (off-board) diagnosis unit 2 and stored there in a second database 5.

The diagnosis unit 2 contains an evaluating expert system 6, which carries out a manual or automatic evaluation of the identified unknown behavior vectors u ( m _j ) * of the database 5 could perform. The assessment can have the consequence that either an error is recognized in a previously known operating mode or that unknown behavior vectors u ( m _j ) * of the second database 5 as new known behavior vectors u ( m _j ) to a then new known operating mode ( m _j ) are stored in the first database 3.

Incidentally, in the first diagnostic unit 1 are the 1 further essentially self-explanatory modules for error detection are shown.

As well as 2 (based on the manageable system complexity) as well as 3 (based on the manageable variety of variants) represent a qualitative differentiation between the proposed methods and the state of the art.

System overview

As in 4th general (and in 1 in detail), according to the invention, for the use of behavior vectors for diagnostic purposes, the system to be monitored is supplemented by a system-related diagnostic unit (on-board unit) and an optional off-board unit. The tasks of the two units and the advantages of the optional off-board unit are explained below.

On-board unit

The task of the on-board unit is to monitor the technical system (here the vehicle) at the sensor level. A diagnostic service is carried out on a control unit ( ECU ) executed. This detects the current operating mode (see explanations in section 3) and assigns a so-called behavior vector (see section 3.1). In addition, the evaluation algorithms for error and anomaly detection as well as for RUL Prediction must be implemented (if the off-board unit is dispensed with).

Off-board unit

The off-board unit consists of a database which contains known behavior vectors (with associated operating modes) and an optional arithmetic unit on which algorithms for anomaly detection and error detection as well as the RUL Forecast are implemented.

Using an off-board unit offers the advantage of enabling fleet-based learning (see Diagnostic Use Case for Detecting Known Errors, Section 3.1.1).

Behavioral vectors for vehicle diagnosis

Behavior vectors represent the dynamic system behavior in vector notation.

The complex, time-discrete system state can be described as the combination of individual, independent sub-states in the high-dimensional state space. A (sub) state is a physical property that can be observed using sensors or indirectly inferred using state estimators.

Motor vehicles are topologically highly complex systems and consequently have a high-dimensional state space. By monitoring this state space, multivariate time series can be observed. In principle, the dynamic system behavior is contained in these time series. However, the time series cannot be used directly for diagnostic purposes, but must be processed.

The concept of behavior vectors is the dynamic expansion of time-discrete representations in the state space. For a given mode of operation m _j (for example an acceleration process: v _-1 = 30 km / h → v ₀ = 50 km / h) time-discrete, complex system states for the current s ₀ (t = 0) and the previous point in time s _-1 (t = -1) can be determined. The behavior vector u is then the connection vector that connects the current state with the previous state (see formula (1)). $$u=s_0-s_{-1}.$$

If a simple, two-dimensional state space is considered, the relationship between successive states and the behavior vector can be graphically illustrated 5 can be removed.

In a sense, behavior vectors describe the path in state space that a system takes during a specific operating mode. Since the representation with behavior vectors is based on physical properties, it does not follow a black box paradigm compared to other system representations. The analysis of behavior vectors thus provides a comprehensible insight into behavior patterns of the system (such as patterns prone to degradation).

Diagnosis use case

Within the diagnosis use case there are three possible applications using behavior vectors. All 3 applications can be assigned to the data-driven methods, but use different learning methods from the field of machine learning ( ML ) as shown in the following table: application Learning process Detection of known errors Supervised ML process Anomaly detection (possible errors) Semi-supervised ML process Discovering unknown errors Semi-supervised ML process

For easier understanding, the methodical approach for all three applications can be based on a two-dimensional state space, as in 6th illustrated, explain. From a mathematical point of view, however, the procedure can be scaled to higher dimensions without any problems.

Detection of known errors

In general, system behavior can be classified as faulty (not tolerated) or faultless. In both cases there are known behavior vectors for known operating modes. In the event of an error, the corresponding behavior vector can be used as a complex error symptom (analogous to current diagnostic systems such as GFS ) can be interpreted.

By monitoring the system (e.g. vehicle in the field) behavior vectors can be collected. In the example in 6th , there are three sets of behavior vectors (u ₁ ... u ₃ ), which the diagnostic system prior are known. The system monitors a code vector u * ₂ which is similar to the known vectors u. ₂ To determine this similarity, ML classification algorithms ( SVM , MLP , etc.) can be used.

• - Die Methode ist in der Lage bedingte Fehler zu erkennen (Fehler, die sich aufgrund eines vorangegangenen Zustands einstellen)
• - Regelbasierte Diagnosefunktionen inklusive eventueller Grenzwerte müssen nicht durch Experten definiert werden. Stattdessen erlernen sich die Funktionen inkl. notwendiger Grenzwerte selbst (unter Verwendung von ML-Methoden/ Ähnlichkeitsmaße)
• - Im Vergleich zu Symptom-basierten Diagnose Systemen (z. B. DTCbasierte Diagnose), nutzen Verhaltensvektoren eine kontinuierliche Skala. Während DTCs erst durch das Verletzen von „Regeln“ gesetzt werden, lassen sich Verhaltensvektoren immer beobachten

The use of behavior vectors for fault detection offers three advantages over known symptom-based approaches:

• - The method is able to recognize conditional errors (errors that occur due to a previous status)
• - Rule-based diagnostic functions including any limit values do not have to be defined by experts. Instead, the functions, including the necessary limit values, learn by themselves (using ML methods / similarity measures)
• - Compared to symptom-based diagnosis systems (e.g. DTC-based diagnosis), behavior vectors use a continuous scale. While DTCs are only set when “rules” are violated, behavior vectors can always be observed

Anomaly detection (possible errors)

The aim of anomaly detection is to identify atypical / unknown (behavior) patterns in the system. While known approaches (e.g. SVM / auto encoder based methods) monitor time-discrete multidimensional states, the method monitors multidimensional system behavior based on behavior vectors. The short-term past is also taken into account.

For example ( 6th ) the observed behavior vector u ₄ * cannot be assigned to any of the known patterns (u ₁ ... u ₃ ). Consequently, the methodology is intended to classify it as an anomaly.

In order to decide whether or not there is an anomaly, similarity measures using the vector properties (direction, position in space, length, etc.) can be applied. From a mathematical point of view, the distances between the observed u * and known u vectors are calculated. The smallest distance measure found must then be compared with a limit value dth. This limit value can either be set manually by experts or automatically determined by a learning algorithm. If the similarity distance measure exceeds the limit value, the observed behavior vector is classified as an anomaly (cf. formula (2)). An anomaly represents a possible failure. For diagnostic purposes, the anomaly still has to be assessed by experts and classified as tolerated behavior or an error. $$a\left(u^{\ast }\right)=\{\begin{array}{cc}1;& dist\left(u^{\ast },u\right)>d_{t\mathrm{<figure-callout\; id="0"\; label="H"\; filenames="DE102020107528A1\_0011.png"\; state="\{\{state\}\}">H</figure-callout>}}\\ 0;& dist\left(u^{\ast },u\right)<d_{tH}\end{array}$$

The main difference (and advantage) of anomaly detection using behavior vectors compared to state-based methods (e.g. SVM-based) is the consideration of the dynamics of the system. While state-based methods only monitor the relevant discrete-time state space, the methodology based on behavior vectors takes the short-term past into account. Faults that occur due to previous states can be easily detected and differentiated from other faults.

Discovering unknown errors

While in theory all possible errors should be covered by diagnostic functions, this cannot be guaranteed in practice. In particular, the networking of sensors, actuators, control units and cross-domain functions presents developers with the challenge of identifying complex faults as early as the development stage. In order to support workshops in identifying the causes of insufficiently secured fault cases, a methodology based on behavior vectors can identify unknown fault cases during the diagnostic session.

Assuming the mode of operation m _j Knowing (including all relevant environmental parameters such as temperature, humidity, road conditions, etc.) can lead to previously unknown errors caused by the Comparison of observed, error-prone behavior vectors u * ( m _j ) with known non-error-prone behavior vectors u ( m _j ) the same operating mode m _j , getting discovered.

In the above example according to 6th the observed behavior vector u ₄ * can be associated with the known behavior vectors u ₄ , provided the same operating mode m _j is present. As in the anomaly detection method, a distance measure is used to classify u ₄ * as an anomaly. Additionally stand with the information about m _j known u ( m _j ) to disposal. From this, the diagnostic system can conclude that for a known operating mode, there is a hitherto unknown system behavior that has not been tolerated. There is thus no need for expert anomaly assessment. Instead, a previously unknown error can be assumed.

Error prognosis and health management ( PHM ) Use case

While the diagnosis use case considers error cases that have already occurred, it is the goal of the PHM Use cases to predict future failure cases. Known approaches use so-called health indicators (HI) to approximate failure probability functions for discrete points in time in the future. In this context, the challenge lies in determining the suitable HIs for the most accurate possible approximation of the failure probability function.

Because behavior vectors u A classified system behavior in a continuous space, using real, physical dimensions, can describe the temporal progressive deviation of a vector u can be used as a continuous indicator of system degradation, as in 7th is shown.

By monitoring behavior vectors u an operating mode m _j Over time, deviations in the system behavior can be detected and quantified. In addition, the time derivative of this deviation can provide an indicator of errors that will "soon" occur. The methodology can predict both known and unknown errors.

The vector properties can be used to quantify deviations in the system behavior. In 7th are two behavior vectors u ( m _j ) and u ^τ ( m _j ) of the same operating mode m _j shown at two different points in time in the two-dimensional state space. The behavior vector u ( m _j ) represents ideal system behavior without signs of degradation (e.g. new vehicle). The behavior ^vector u τ ( m _j ) is observed at a later point in time τ after progressive degradation of the system.

• - Überwachung des Winkels δ^τ zwischen u(m_j ) und u^τ(m_j )
• - Überwachung der Lage von u(m_j ) im Zustandsraum
• - Überwachung der Länge von u(m_j )
• - Eine Kombination der obigen Methoden

To for the PHM Use case to quantify deviations from u, one or the following four methods can be applied:

• - Monitoring the angle δ ^τ between u ( m _j ) and u ^τ ( m _j )
• - Monitoring the position of u ( m _j ) in the state space
• - Monitoring the length of u ( m _j )
• - A combination of the above methods

All of the methods described follow a white box paradigm and are therefore understandable for system experts.

Use case quality assurance ( QA )

The causes of errors in technical systems are manifold. In particular, the causes of the errors that occur during the warranty period are of great interest for quality assurance. While tester systems are mainly used to determine the smallest exchangeable units in the workshops, the diagnosis is aimed at QA on the identification of the fundamental cause of the error on a physical level. the end QA Therefore, there are different requirements for a diagnostic system.

A methodology suitable for quality assurance must collect error data in uncompressed form so that the basic physical processes that led to the error can be traced. This enables error cases to be reproduced under laboratory conditions. Behavioral vectors represent precisely this information (provided that it was recorded in the event of an error).

An error case is relevant for quality assurance if it is observed more often than would be statistically expected. It can therefore be assumed that a sufficient amount of case data is known. Since the operating mode is known from the case data in addition to the observed behavior vectors (of the error case), “healthy” behavior vectors of the operating mode are also available.

In 8th a two-dimensional state space is shown again with a simple example. They are for one operating mode m _j an observed behavior vector u ^o ( m _j ) as well as the expected behavior vector u ^e ( m _j ). With the help of these vectors the behavior vector of the error u ^f ( m _j ) reconstruct (see formula (3)). It represents the deviation between target and actual behavior. Furthermore, the amounts of the individual vector entries of u ^f ( m _j ) with the importance of the corresponding partial states in the event of an error. In a way, u ^f ( m _j ) can be seen as a sensitivity vector, which maps the error sensitivity to the multidimensional state. $$u^f\left(m_j\right)=u^O\left(m_j\right)-u^e\left(m_j\right)$$

In addition to the promising failure-specific analysis of the u ^f ( m _j ) Vectors, we hope to find general, error-case-independent patterns in the u ^f . Directions or amounts of u ^f could, for example, indicate electrical, communication-related, mechanical or other faults. This would result in options using machine learning to automatically carry out a general error assessment (for example, to make action decisions).

List of reference symbols

CM

Condition Monitoring

DTC

Diagnostic Trouble Code

ECU

Electronic Control Unit

GFS

Geführte Fehlersuche

Hl

Health Indicator

ML

Maschinelles Lernen

MLP

Multi-Layer Perceptron (Verfahren des Maschinellen Lernens)

PHM

Prognostics and Health Management

QA

Quality Assurance

RUL

Remaining Useful Lifetime

SVM

Support Vector Machine (Verfahren des Maschinellen Lernens)

Symbolverzeichnis

α(u)

Als Anomalie klassifizierter Verhaltensvektor u

dth

Distanz Grenzwert (als Ähnlichkeitsmaß)

fi

Dimension i (Physikalische Eigenschaft) im komplexen Zustandsraum

mj

Eindeutiger Betriebsmodusj

si

Eindeutiger Zustand zum Zeitpunkt i (Vektor)

u, u*

Verhaltensvektor, Beobachteter Verhaltensvektor

CM

Condition monitoring

DTC

Diagnostic Trouble Code

ECU

Electronic control unit

GFS

Guided troubleshooting

St.

Health indicator

ML

Machine learning

MLP

Multi-Layer Perceptron (machine learning method)

PHM

Prognostics and Health Management

QA

Quality assurance

RUL

Remaining Useful Lifetime

SVM

Support Vector Machine (method of machine learning)

Symbol directory

α (u)

Behavior vector classified as an anomaly, etc.

dth

Distance limit value (as a measure of similarity)

fi

Dimension i (physical property) in the complex state space

mj

Unique operating mode j

si

Unique state at time i (vector)

u, u *

Behavior vector, observed behavior vector

Claims (3)

Diagnostic system for anomaly detection and / or error prognosis in motor vehicles, which classifies known errors as well as unknown errors on the basis of behavior vectors (u), with an internal vehicle diagnostic unit (on-board unit, 1) and / or with a diagnostic unit (off-board unit, 2) external to the motor vehicle, at least one of the two diagnostic units (1, 2) containing a first database (3) in which the known operating modes (m _j ) are known Behavioral vectors (u (m _j )) are stored, and in at least one of the two diagnosis units (1, 2) a diagnosis Control unit (4) with an evaluation algorithm for anomaly detection and / or error prognosis is included, which has a vector similarity test for comparing determined behavior vectors with known behavior vectors (u (m _j )) of the database (3) in order to identify unknown behavior vectors (u (m _j ) *) to identify. Diagnostic system according to Claim 1 , characterized in that identified unknown behavior vectors (u (m _j ) *) are stored in a second database (5) in order to use an evaluating expert system (6) either to detect errors in a previously known operating mode or as new known behavior vectors (u (m _j )) for a then new known operating mode (m _j ) in the first database (3). Diagnostic system according to one of the preceding claims, characterized in that the diagnostic control unit (4) is further designed such that a failure probability is predictable after a tendency of a behavior vector (u) to deviate from an operating mode (m _j ) over time with respect to angle, Position and / or length was detected and quantified.

Images