DE102021210106A1 - Computer-implemented method and system for anomaly detection and method for anomaly detection in a final acoustic test of a transmission - Google Patents

Abstract

Computer-implemented method for anomaly detection comprising the steps of: Obtaining data from N states (Z) of a technical object, the data being measured by sensors and/or generated in a simulation of the technical object and/or calculated from post-processing of data measured by sensors became (C1); ordering the data of the respective states by h x k positions, each entry of one of the N states at one of the h x k positions being comparable to the entry of another of the N states at the same position, and obtaining a training data set in the form of an N × h × k arrangement; Training an analysis model (ANOM-Stat) by feeding the training states into the analysis model, the analysis model determining a statistical characterization of individual positions with regard to anomalies by calculating a reference state using l reference images, the reference images each being divided into h ■ k position images gi,j; Feeding a state to be checked into the trained analysis model (ANOM-Stat) (C4), the trained analysis model (ANOM-Stat) comparing the state to be checked with the reference state by means of at least one difference mapping and obtaining at least one difference state (C5), the difference state assigns an anomaly value (AS) according to a figure (C6)

Description

The invention relates to a computer-implemented method and a system for anomaly detection. The invention also relates to a method for detecting anomalies in a final acoustic test of a transmission.

Industrial manufacturing processes are variable and complex processes that are correspondingly prone to errors and failures. At the same time, as part of the so-called 4th industrial revolution, the penetration of manufacturing processes with sensors is increasing, which ensures that the processes are comprehensively mapped by data. Ensuring that an industrial process step runs smoothly is associated with a great deal of effort and high costs, since error detection and error elimination are often manual processes and also take place reactively. In this context, anomaly detection is a fundamental challenge in industrial processes.

The object of the invention was how, based on status data, a degree of an abnormality of the status can be mapped.

The objects of claims 1, 9 and 10 solve this problem.

• • Erhalten von Daten von N Zuständen eines technischen Objekts, wobei die Daten mittels Sensoren gemessen und/oder in einer Simulation des technischen Objekts erzeugt und/oder aus einer Nachverarbeitung von mittels Sensoren gemessenen Daten errechnet wurden;
• • Beschreiben jeden der N Zustände durch eine mathematische Anordnung gleicher Dimensionalität;
• • Ordnen der Daten der jeweiligen Zustände nach h · k Positionen, wobei jeder Eintrag eines der N Zustände an einer der h · k Positionen mit dem Eintrag eines anderen der N Zustände an derselben Position vergleichbar ist und Erhalten eines Trainingsdatensatzes in Form einer N × h × k Anordnung;
• • Trainieren eines Analysemodells durch
  • ◯ Einspeisen der Trainingszustände in das Analysemodell, wobei
  • ◯ das Analysemodell eine statistische Charakterisierung einzelner Positionen hinsichtlich Anomalien bestimmt durch Berechnung eines Referenzzustand x_ref ∈ R^h×k×l mittels l Referenzabbildungen ƒ: R^N×h×k → R^h×k, wobei die Referenzabbildungen jeweils in h · k Positionsabbildungen g_i,j: R^N → R; i ≤ h, j ≤ k unterteilt sind, ein Wert einer Position im Referenzzustand x_ref ∈ R^h×k×l gemäß den Positionsabbildungen g_i,j aus den Werten der Position in den Trainingszuständen berechnet wird und
  • ◯ Erhalten eines trainierten Analysemodells;
• • Einspeisen eines zu überprüfenden Zustandes x_test ∈ R^h×k in das trainierte Analysemodell, wobei das trainierte Analysemodell
  • ◯ den zu überprüfenden Zustand x_test mit dem Referenzzustand x_ref mittels wenigstens einer Differenzabbildung d: R^h×k × R^h×k×l → R^h×k vergleicht und wenigstens einen Differenzzustand x_diff = d(x_test,x_ref) erhält,
  • ◯ dem Differenzzustand x_diff gemäß einer Abbildung f_score: R^h×k → R einen Anomaliewert zuordnet;
• • Erkennen der Anomalien basierend auf dem Anomaliewert.

In one aspect, the invention provides a computer-implemented method for anomaly detection. The procedure includes the steps:

• • Obtaining data from N states of a technical object, the data being measured using sensors and/or generated in a simulation of the technical object and/or being calculated from post-processing of data measured using sensors;
• • Describe each of the N states by a mathematical arrangement of equal dimensionality;
• • ordering the data of the respective states by h x k positions, where each entry of one of the N states at one of the h x k positions is comparable to the entry of another of the N states at the same position and obtaining a training data set in the form of an N x h × k arrangement;
• • Train an analysis model
  • ◯ Feeding the training states into the analysis model, where
  • ◯ the analysis model determines a statistical characterization of individual positions with regard to anomalies by calculating a reference state x _ref ∈ R ^h×k×l using l reference maps ƒ: ^RN×h×k → R ^h×k , where the reference maps are each in h · k position maps g _i,j : R ^N → R; i ≤ h, j ≤ k, a value of a position in the reference state x _ref ∈ R ^h×k×l is calculated according to the position maps g _i,j from the values of the position in the training states, and
  • ◯ obtaining a trained analysis model;
• • Feeding a state to be checked x _test ∈ R ^h×k into the trained analysis model, with the trained analysis model
  • ◯ compares the state to be checked x _test with the reference state x _ref by means of at least one difference mapping d: R ^h×k × R ^h×k×l → R ^h×k and at least one difference state x _diff = d(x _test ,x _ref ) receives,
  • ◯ assigns an anomaly value to the difference state x _diff according to a mapping f _score : R ^h×k → R;
• • Detecting the anomalies based on the anomaly value.

The mathematical arrangement is, for example, an h×k matrix. Then every state is 2-dimensional. The mathematical arrangement can also be an h×k×l matrix. In this case, each state is 3-dimensional. The dimension of the state is irrelevant for the invention described here; what is important is that all states considered have the same dimension and the entries (positions) of the state can be sorted. The invention is described using an h×k matrix in order to make it easier to understand and read, without restricting the generality. For example, this could be a sonagram from a transmission acoustic test. Then each state corresponds to a time/frequency coordinate.

In one aspect, the invention also provides a computer program for anomaly detection. The computer program includes instructions that cause a computer to carry out the steps of the inventive method for anomaly detection when the computer program runs on the computer.

In one aspect, the reference maps, the position maps, and/or the anomaly value map are determined using hyperparameter optimization.

According to a further aspect, the reference state is calculated using 2 reference images, each with a position image. A first position map is an averaging function. A second position mapping is a standard deviation. Alternatively, the reference state by means of a reference image with a Positionsab education calculated. The position mapping is a quantile function.

According to a further aspect, the difference state x _diff is calculated according to x _diff = x _test - (x _mean + 3 * x _std ). x _mean and x _std are the parts of the reference state x _ref that result from the calculation of the mean and the standard deviation over the training states. The entries in the difference state that are greater than zero then stand for positions at which a value in the state to be tested lies at the edge of the distribution of the corresponding values in the training data, namely above three standard deviations from the mean. An equivalent procedure can be performed for very small values, i.e. subtracting the mean minus three standard deviations.

According to a further aspect, m difference states are calculated using m difference images. The anomaly value is assigned to the m difference states according to a mapping f _score : R ^h×k×m →R.

How this difference is calculated depends on the type of reference and the respective application. It is crucial that the l references and the state to be checked have the same dimension and are ordered in the same way, so that a position-by-position comparison can be carried out in the form of at least one difference mapping.

According to a further aspect, the anomaly value is calculated as the mean of all positive values in the difference state or states x _diff . This corresponds to a global averaging pooling operation. In general, any operation that maps the difference state to a scalar value can be used, for example pooling operation from convolutional networks, for example min/max pooling. According to one aspect of the invention, the anomaly value is calculated from the difference state using a convolution network.

The way in which the anomaly value is calculated from the difference state is in turn dependent on the application and can be determined in a data-driven manner, for example by means of hyperparameter tuning.

In a further aspect, the hxk positions include coordinates in a 2-dimensional space, e.g., time and frequency. This allows measured values to be compared that belong to the same time and the same frequency of a series of measurements.

According to a further aspect, the training states and/or the state to be checked are x _test sonagrams. This allows noise data to be analyzed in the frequency domain. The noise signals are transformed into the frequency space, for example by means of Fourier transformation or wavelet transformation. This is how the sonagrams are created. In a sonagram, an entry corresponds to the amplitude or energy of a frequency in a time interval, i.e. a matrix of numerical values results in which each column corresponds to a time interval and each row to a frequency. The entries of a sonogram are thus ordered.

• • ein erstes Modul, das Daten von Zuständen eines Prozesses, eines Bauteils und/oder einer Produktionsmaschine in wenigstens einem zu überwachenden Prozessschritt des industriellen Fertigungsprozesses zu einem bestimmten Zeitpunkt oder in einem bestimmten Zeitintervall, aus gleichartigen Prozessschritten und/oder aus nachgelagerten Prozessen erhält, umfassend sensoriell gemessene Daten und/oder aus Simulieren des Prozesses erhaltene Daten, die Daten verarbeitet und ein Analysemodell nach einem der voran gehenden Ansprüche trainiert und erhält;
• • ein zweites Modul, dem das trainierte Analysemodell bereitgestellt wird und das einen Anomaliewert berechnet;
• • ein drittes Modul, das den Anomaliewert einer Überprüfungsinstanz bereitstellt, die,
  • ◯ falls der erhaltene Anomaliewert einen anormalen Zustand klassifiziert mit hoher Genauigkeit, den Zustand als Anomalie ohne weitere Überprüfung meldet;
  • ◯ falls der erhaltene Anomaliewert einen anormalen Zustand klassifiziert, eine Ähnlichkeit zu einem bereits als potentielle Anomalie erkanntem und annotiertem Zustand bestimmt,
  • ◯ im Falle hoher Ähnlichkeit eine Annotation des als Anomalie erkannten Zustandes übernimmt,
  • ◯ andernfalls in Abhängigkeit einer Größe des erhaltenen Anomaliewertes den Zustand annotiert.
  • ◯ einen als normal klassifizierten Zustand mit gewisser Wahrscheinlichkeit annotiert.

According to a further aspect, the invention provides a system for anomaly detection in industrial manufacturing processes. The system includes

• • a first module that receives data from states of a process, a component and/or a production machine in at least one process step of the industrial manufacturing process to be monitored at a specific point in time or in a specific time interval, from similar process steps and/or from downstream processes sensorially measured data and/or data obtained from simulating the process, processing data and training and maintaining an analysis model according to any one of the preceding claims;
• • a second module that is provided with the trained analysis model and that calculates an anomaly score;
• • a third module that provides the anomaly value to a verification instance that,
  • ◯ if the obtained anomaly score classifies an abnormal condition with high accuracy, reports the condition as an anomaly without further verification;
  • ◯ if the obtained anomaly value classifies an abnormal state, determines a similarity to a state that has already been recognized and annotated as a potential anomaly,
  • ◯ in the case of high similarity, takes on an annotation of the state recognized as an anomaly,
  • ◯ otherwise annotates the state depending on a size of the received anomaly value.
  • ◯ annotates a condition classified as normal with a certain probability.

The system according to the invention integrates the following steps and/or modules: data-based definition of the status of the process to be monitored, data storage and processing of all necessary data, training for an anomaly detection Analysis based on historical data, integration of a verification instance, e.g. a human verifier, to control the detected anomalies and to generate further annotations for the targeted expansion of the training database, detection of anomalies, processing of the evaluation of the human verifier and based on this continuous adjustment and improvement of the anomaly detection . The basic functionalities, such as saving and expanding the database and/or requesting annotations to expand the database, are taken over by the first module. The task of detecting potential anomalies is performed by the second module. In this case, the process or production machine to be monitored continuously sends data or test results, corresponding to the defined state Z, to the system according to the invention. The system response then contains a classification of the state Z as "normal/abnormal", optionally with a confidence estimate, and is sent to a human expert for verification via the third module.

• • Antreiben des Getriebes mit einem Drehzahlprofil, Messen des entstehenden Körperschalls mittels Akustiksensoren und Erhalten von Körperschall-Zeitreihen;
• • Transformieren der Körperschall-Zeitreihen in einen Frequenzraum und Erhalten von Sonagrammen;
• • erfindungsgemäßes Trainieren eines Analysemodells, wobei die Trainingszustände historische Sonagramme von in Ordnung und nicht-in-Ordnung getesteten Getrieben umfassen und die Auswahl von Referenzabbildungen, Positionsabbildungen, Differenzabbildungen und/oder Abbildungen für einen Anomaliewert des Analysemodells während des Trainings Daten getrieben erfolgt.

According to a further aspect, the invention provides a method for anomaly detection in a final acoustic test of a transmission. The procedure includes the steps

• • Driving the transmission with a speed profile, measuring the resulting structure-borne noise using acoustic sensors and obtaining structure-borne noise time series;
• • transforming the structure-borne noise time series into a frequency space and obtaining sonagrams;
• • inventive training of an analysis model, wherein the training states include historical sonagrams of tested ok and bad transmissions and the selection of reference maps, position maps, difference maps and/or maps for an anomaly value of the analysis model during the training is data-driven.

After the completion of a gearbox, the flawless condition is examined, among other things, by means of an acoustic test. The gearbox is driven with a standardized speed profile and the resulting structure-borne noise is measured. The speed profile consists, among other things, of speed ramps in different gears. Anomaly detection according to the invention is carried out, optionally in addition to an established check for conspicuous noises. The structure-borne noise time series that arises for each transmission test is transformed into the frequency domain using Fast Fourier Transformation or Wavelet Transformation. This produces a sonagram that satisfies the above definition of ordered numeric data. In this application, the sonagrams of gearboxes (of the same type) tested as “okay” (OK) from the past are used as training data. A suitable separation of the existing data into training, validation and test data is assumed here as a standard procedure and is therefore not mentioned explicitly. The data-driven selection of the necessary mappings of the analysis model according to the invention is carried out using the available historical "not OK" (NOK) data. This means that the images are selected in such a way that the resulting analysis model can distinguish NOK gears from OK gears as best as possible.

In one aspect, the anomaly detection method is computer implemented in a final acoustic test of a transmission.

According to one aspect, the invention also provides a computer program for anomaly detection in a final acoustic test of a transmission. The computer program includes instructions that cause a computer to carry out the steps of the method according to the invention for anomaly detection in a final acoustic test of a transmission when the computer program runs on the computer.

• • die Referenzabbildung der Berechnung des 99%-igen Quantils pro Position;
• • die Differenzabbildung der Berechnung der Differenz zwischen einem zu überprüfenden Zustand x_test und einem Referenzzustand x_ref; und/oder
• • die Abbildung für den Anomaliewert der Berechnung des Mittelwerts aller positiven Werte in dem erhaltenen Differenzzustand x_diff.

According to one aspect, the method according to the invention for anomaly detection corresponds to an acoustic final inspection of a transmission

• • the reference figure of the calculation of the 99% percentile per position;
• • the difference mapping of the calculation of the difference between a state to be checked x _test and a reference state x _ref ; and or
• • the mapping for the anomaly value of the calculation of the mean of all positive values in the obtained difference state x _diff .

On the way to achieving the solution according to the invention, the following aspects have surprisingly been shown: Inaccuracies in the switching process can lead to increased noise in the structure-borne sound signal in the corresponding areas. The 99% quantile is far enough towards the upper end of the distributions that these influences are suppressed. In the case of the transmission test, "too loud" frequencies/anomalies are of particular interest; simply calculating the difference allows these values to appear as "all values greater than zero". In the case of the 99% quantile, all values in the difference state have an entry greater than zero that (in their respective position) in the initial state, i.e. the one to be checked state, have a value greater than 99% of the entries in the training data. Since, in addition to the number of frequencies that are too loud, the level of the deviations also plays a role, the mean value of the deviations is an advantageous parameter for characterizing the deviations in a single value.

According to a further aspect, a system according to the invention is used to carry out the method for anomaly detection in a final acoustic test of a transmission.

Advantageous configurations of the invention result from the following definitions, drawings and the description of preferred exemplary embodiments.

The technical object can be a component, a production machine or a process step in an industrial manufacturing process, for example a gearbox.

• • Bauteil-bezogene Sensormessung: Geometrie, Aussehen;
• • Maschinen-bezogene Messung: Vibrationen und Temperaturen einer oder mehrerer Komponenten;
• • Sensor in einem Bauteiltest: Geräusch, Vibration, Stromstärke, Druck.

In order to acquire data and process it further, the invention makes use of sensors and/or sensor models that are used, for example, in industrial processes. Sensor models simulate real sensors. The term sensor includes real sensors and sensor models. One way to categorize these sensors is based on their physical measurand, e.g. temperature sensors, vibration sensors, force and pressure sensors, optical sensors including camera, infrared, lidar and radar sensors, measuring the size of a component including its geometry, current and voltage sensors and others. Which sensors are used in a specific case depends on the respective process step. According to one aspect of the invention, functions of a machine and also the properties of a component are monitored by means of sensors, including the aforementioned sensors. According to one aspect, the invention includes the following sensor configurations:

• • Component-related sensor measurement: geometry, appearance;
• • Machine-related measurement: vibrations and temperatures of one or more components;
• • Sensor in a component test: noise, vibration, current, pressure.

According to one aspect, the sensors use Internet of Things technology to transmit data, for example to one another, to individual components/modules of the solution according to the invention and/or to a cloud infrastructure. Automated or autonomous anomaly detection can thus be implemented. The cloud infrastructure includes cloud-based data storage. The cloud infrastructure is, for example, a public, private or hybrid cloud infrastructure.

According to one aspect of the invention, the definition of what an anomaly is can be based on the definition of what is to be understood by a state in the industrial manufacturing process.

Part of the solution according to the invention is based on the definition of the status of the production step or component to be monitored with regard to the data describing the status. The status definition can form the basis for further steps up to the detection of anomalies by defining the data basis for the training of models and the use of the solution according to the invention. The definition of both a state and the associated anomalies are not static concepts, but can be subject to change over time, for example because new sensors are available, quality requirements are changed, or materials involved in the process or product are changed.

• • Der Zustand einer Maschine kann durch Gesamtheit der Sensordaten innerhalb eines Zeitintervalls, beispielsweise der letzten 10 Sekunden, definiert sein, sowie zusätzlich durch Bauteil-spezifische Parameter.
• • Der Zustand eines Bauteils kann durch die Ergebnisse eines Tests definiert sein, sowie durch weitere Bauteilparameter, umfassend unterschiedliche Bauteil-Varianten.

In Abhängigkeit der jeweiligen Zustandsdefinition wird eine Verteilung der Zustände bestimmt.Condition describes the condition/properties of a process, a component and/or a (production) machine at a specific point in time or in a specific time interval. This state is detected by measurements using suitable sensors as described above. Which sensors are suitable depends on the definition of the state as well as the system to be described (component, machine, ...). Conversely, the availability of appropriate sensors also influences the definition of the state. A state that is not accessible by appropriate sensors is not a meaningful definition of a state. A state can have different characteristics in the industrial manufacturing process, the invention includes the following state definitions:

• • The state of a machine can be defined by all of the sensor data within a time interval, for example the last 10 seconds, and also by component-specific parameters.
• • The condition of a component can be defined by the results of a test, as well as by other component parameters, including different component variants.

A distribution of the states is determined depending on the respective state definition.

An anomaly is a condition, specified by corresponding (sensor) data, which both is rare as well as differing from almost all other conditions based on the recorded data. An anomaly is rare and different. Nevertheless, it can often be difficult to draw a precise distinction between normal and abnormal, and in many cases even not possible with the available data, so that ultimately a condition is only abnormal with a certain probability. In order to do justice to this circumstance, a checking authority, for example a human checker, is proposed as part of the system according to the invention.

Ordered numerical data forms the data basis of the solution according to the invention. Ordered means that the individual entries in the data that describe a state can be arranged in an order so that each entry in one state can be meaningfully compared with the entry in another state at the same position. The states are compared position by position.

The solution according to the invention can evaluate the state of an industrial production step and/or the ordered data associated with a state by statistical characterization of individual positions with regard to its anomaly. In a first step, an analysis model is trained. A statistical characterization is calculated with the help of training data. The training data includes a large number of historical states of the process step to be monitored, which if possible contain no abnormal states - in reality at least few. Instead of historical data of the actual process, data from comparable processes or from a simulation of the process can also be used.

The training manifests itself in a quantitative assessment of the distribution of the training states (almost only normal states) at the position level, this so-called reference can be calculated in different ways.

Let X be a set of N training states X = (x ₁ , x ₂ ,...,x _N ), x _L ER ^h×k _; h, k ∈ N*, X ∈ R ^N×h×k, i.e. each training state is described by an h × k matrix of ordered numerical values (in general, the states can also have more than two dimensions, the description of the two-dimensional case is chosen here for the sake of readability and clarity, without restricting generality). Then the calculation of the reference consists of one or more mappings ƒ: ^RN×h×k → R ^h×k , called reference mappings in the following. This means that the reference x _ref has the dimension Rh ^×k×l , where l is the number of reference images.

Each of the reference mappings is in turn subdivided into h · k further mappings g _i,j : ^RN → R, i≦h, j≦k, which are referred to below as position mappings. These position mappings establish the calculation rule with which the value of a position in the reference state is calculated from the values of the position in the training states. In the simplest case, the same mapping can be used for each position, ie g _1,1 = g _1,2 = . . . = g _h,k = g or, in the other extreme case, a different mapping g _1,1 for each position ≠ g _1.2 ≠ ... ≠ g _h,k can be used.

In one aspect, more complex position mappings are used, for example, using a window approach, multiple positions in the training data are used to calculate the value of a position in the reference state.

After the analysis model has been trained, the analysis model is applied in a second step. In this case, an unknown state, ie the ordered numerical data describing a state, is evaluated by the trained analysis model with regard to its anomaly. This score is passed on in the form of an anomaly score.

After the training, statuses originating from a production step whose data formed the basis of the training, for example sonagrams during the final inspection of gearboxes, can be checked for anomalies. The analysis model calculates an anomaly value that measures the degree of deviation from the expected, normal behavior of the production step. In general, the higher the anomaly score, the greater the deviation.

The modules of the system according to the invention include hardware and/or software modules, including hardware and/or software modules for regulating and/or controlling industrial manufacturing processes and/or for anomaly detection. The hardware modules include electronic units, integrated circuits, embedded systems, microcontrollers, multiprocessor systems-on-chip, central processors and/or hardware accelerators, e.g. graphics processors, data storage units and connectivity elements, e.g. WLAN modules, RFID modules, Bluetooth modules, NFC modules . According to one aspect of the invention, the anomaly detection is implemented as functional software in the cloud infrastructure.

The instructions of the computer programs according to the invention include machine instructions, source text or object code written in assembler language che, an object-oriented programming language, such as C++, or in a procedural programming language, such as C. According to one aspect of the invention, the computer programs are hardware-independent application programs that are provided, for example, via a data carrier or a data carrier signal using software over the air technology.

• 1 ein Ausführungsbeispiel eines erfindungsgemäßen System,
• 2 ein Ausführungsbeispiel der erfindungsgemäßen Anomalieerkennung im Kontext des erfindungsgemäßen Systems,
• 3 eine schematische Darstellung geordneter numerischer Zustände,
• 4 eine Darstellung zu Trainings- und Anwendungsphase des erfindungsgemäßen Verfahrens zur Anomalieerkennung,
• 5 eine Darstellung einer erfindungsgemäßen Trainingsphase,
• 6 eine Darstellung einer erfindungsgemäßen Anwendungsphase,
• 7 ein Ausführungsbeispiel eines Sonagramms,
• 8 ein Ausführungsbeispiels des erfindungsgemäßen Verfahrens zur Anomalieerkennung und
• 9 ein Ausführungsbeispiel eines erfindungsgemäßen Verfahrens zur Anomalieerkennung in einer akustischen Endprüfung eines Getriebes.

The invention is illustrated in the following exemplary embodiments. Show it:

• 1 an embodiment of a system according to the invention,
• 2 an exemplary embodiment of the anomaly detection according to the invention in the context of the system according to the invention,
• 3 a schematic representation of ordered numerical states,
• 4 an illustration of the training and application phase of the method according to the invention for anomaly detection,
• 5 a representation of a training phase according to the invention,
• 6 a representation of an application phase according to the invention,
• 7 an embodiment of a sonagram,
• 8th an embodiment of the inventive method for anomaly detection and
• 9 an embodiment of a method according to the invention for anomaly detection in a final acoustic test of a transmission.

In the figures, the same reference symbols denote the same or functionally similar reference parts. For the sake of clarity, only the relevant reference parts are highlighted in the individual figures.

In 1 the system IF-Anom monitors a production step PS and/or the respectively associated production machine PM with regard to any anomalies that occur, for example a production step PS in the manufacture of a transmission. A first module IF-Anom Core trains an analysis model for anomaly detection based on historical process data Data1, state definitions Z and existing state data Data2. The training data can also include third synthetic data Data3.

The analysis models are trained, for example, at the level of transmission variants and production lines. There is one version of each model for each transmission variant and production line. This restriction is always recorded in the associated configuration Config-Model, so that the model associated with the transmission in question is always loaded later on.

The first module IF-Anom Core includes two essential functions of the IF-Anom system, on the one hand the processing and data storage of incoming data Data1, Data2 using a data processing module, on the other hand the training and evaluation of analysis models for anomaly detection using a training module. Both functions are subject to a time aspect, ie at the beginning of the use of the IF-Anom system, existing data Data2 from the process to be monitored is first accessed and/or data from comparable processes or synthetically generated data. Based on this, the state definition Z is developed. Data and state definition are brought into a form suitable for the subsequent steps and stored.

The second module IF-Anom anomaly detector takes on the task of processing incoming data comprehensively Data2 according to the existing status definition and data pre-processing routines. The incoming data is processed in the same way as the historical data Data1 used for model training. An assessment is then carried out using the available anomaly detection models. The output of each model includes an anomaly score AS. The anomaly value AS indicates how abnormal the evaluated state Z is, usually but not necessarily, the higher the more abnormal. Which methodology and which models are used depends on the application and the availability of models of sufficient quality. The second module IF-Anom anomaly detector receives the current models from the first module IF-Anom Core, as well as an application-specific configuration Config-Model, in which, for example, the necessary pre-processing of the incoming data Data2 is specified. Using these models, the current state Z is then assessed by the second module IF-Anom anomaly detector.

The first module IF-Anom Core and the second module IF-Anomalyidentifier exchange data and/or evaluations with each other.

The assessment of the second module, IF anomaly detector, is passed on to the third module, IF anom annotation manager. The third module IF-Anom annotation manager takes care of that Task to decide whether and when a state Z is given to one, or to which, human expert for assessment. For example, roughly speaking, detected anomalies are almost always propagated, and additionally detected-normal samples under certain circumstances. The evaluations by the human expert go back to the third module IF-Anom Annotation Manager, which decides in what form annotations go back to the first module IF-Anom Core in order to expand the database there accordingly. In the ideal case, the third module IF-Anom annotation manager has access to meta information regarding the assessment, for example an anonymous identification of the assessing expert, especially if necessary for data protection reasons, as well as a time stamp when the assessment was written.

If an NOK anomaly is confirmed by the expert reviewer as an NOK anomaly or if the detected NOK anomaly is released without checking, two reactions are set in motion, for example. On the one hand, the monitored production step PS is examined more closely based on the state Z recognized as abnormal, in order to decide whether/when maintenance/repair must be carried out. On the other hand, the components that were processed during and after the occurrence of the NOK anomaly can be sorted out as a precaution and/or examined separately.

The training and application of the analysis model ANOM-Stat according to the invention is carried out by the first module IF-Anom Core and the second module IF-Anomaly Detector, see also 2 .

3 shows an example of ordered numerical states. The entries in the same position in different states correspond to one another and can be meaningfully compared with one another. For N states, where each state corresponds to an h×k matrix, an N×h×k array is obtained.

4 shows the method according to the invention for anomaly detection divided into a first sub-method for training the analysis model ANOM-Stat and a second sub-method for using or applying the analysis model ANOM-Stat trained according to the first sub-method. The full procedure, including the first and second sub-procedures, is in 8th shown.

The first sub-process can include process steps C1-C3 and is described in detail in 5 shown. In method step C1, data from n states Z, Z1-Z3 of a technical object, for example a transmission, is obtained. The data is measured, for example, using sensors, for example using acoustic sensors, for example microphones. In a method step C2, the data of the respective states after h·k positions, for example after h time intervals and k frequencies, according to 3 orderly. The previous steps are repeated and N training states are obtained as training data X=(x ₁ , x ₂ ,...,x _N ), x _i ∈ R ^{h ×k} . The analysis model ANOM-Stat is trained in a method step C3. The training states are fed into the analysis model ANOM-Stat. The analysis model ANOM-Stat determines a statistical characterization of individual positions with regard to anomalies by calculating a reference state x _ref using I reference images f. A value of a position in the reference state x _ref is calculated according to position images g _i,j from the values of the position in the training states and the result is a trained analysis model.

The second sub-process can include process steps C4-C7 and is detailed in 6 shown. In method step C4, a test state x _test is fed into the trained analysis model ANOM-Stat. In a method step C5, the state x _test to be checked is compared with the reference state x _ref by means of at least one differential image d and at least one differential state x _diff is obtained. In a method step C6, the anomaly value AS is assigned to the differential state x _diff according to a mapping f _score . In a method step C7, an anomaly NOK is detected based on the anomaly value AS.

7 shows a sonagram as an example of ordered numerical states. The sonagram is obtained from a transformation of a time signal, for example an acoustic signal over time, into frequency space. The sonagram corresponds to an ordered sequence of numerical values, i.e. an h × k arrangement. Each column t1, t2, t3, ... represents a time interval. Each row f1, f2, f3, ... represents a frequency. The entries are usually amplitudes or energies of a frequency in a time interval. This sorts the entries.

9 shows the procedure for anomaly detection in a final acoustic test of a transmission. In a method step E1, the transmission is driven with a speed profile, for example with a standardized speed profile. The resulting structure-borne noise is measured using acoustic sensors, for example microphones, and structure-borne noise time series are obtained. In a method step E2, the structure-borne noise time series are transformed into a frequency space and sonagrams, for example as in 7 shown received. In a method step E3, the analysis model ANOM-Stat trai ned. The training states include historical sonagrams of transmissions tested as good. The selection of reference images, position images, difference images and/or images for the anomaly value AS of the analysis model ANOM-Stat is data-driven during the training. In a method step E4, an anomaly is detected based on the anomaly value AS.

Reference List

xref

reference state

xi

training condition

X

training data

f

reference figure

gi, j

position mapping

xtest

condition to be checked/test condition

xdiff

difference state

i.e

difference mapping

fscore

Figure for anomaly score

i,j

position

IF

industrial manufacturing process

hp

process step

p.m

production machine

ANOM Stat

analysis model

IF-Anom

Industrial anomaly detection system

IF-Anom

Core first module

IF-Anom

Anomaly detector second module

IF-Anom

Annotation manager third module

IF-Anom

Data fourth module

Expert

monitor

Data1

first data

Data2

second data

Data3

third data

AS

Anomaly Value/Anomalies Score

NOK

anomaly

Z, Z1-Z3

Condition

config model

Configuration of trained models

t1-t3

Points in time/time intervals

f1-f3

frequencies

C1-C7

process steps

E1-E4

process steps

Claims (12)

Computer-implemented method for anomaly detection comprising the steps • Obtaining data from N states (Z, Z1-Z3) of a technical object, the data being measured using sensors and/or generated in a simulation of the technical object and/or from post-processing using sensors measured data were calculated (C1); • Describe each of the N states (Z, Z1-Z3) by a mathematical arrangement of the same dimensionality; • ordering the data of the respective states by hxk positions, where each entry of one of the N states (Z) at one of the hxk positions is comparable to the entry of another of the N states (Z) at the same position and obtaining a training data set in the form of an N × h × k array; • Training an analysis model (ANOM-Stat) by ◯ feeding the training states into the analysis model (ANOM-Stat), where ◯ the analysis model (ANOM-Stat) determines a statistical characterization of individual positions with regard to anomalies by calculating a reference state x _ref ∈ R ^{h× k×l} by means of l reference mappings ƒ: R ^N×h×k → R ^h×k , where the reference mappings are in h · k position mappings g _i,j : R ^N → R; i ≤ h, j ≤ k, calculating a value of a position in the reference state x _ref ∈ R ^h×k×l according to the position maps g _i,j from the values of the position in the training states (C3) and ◯ obtaining a trained one analysis model; • Feeding a state to be checked x _test ∈ R ^h×l into the trained analysis model (ANOM-Stat) (C4), the trained analysis model (ANOM-Stat) ◯ the state to be checked x _test with the reference state x _ref using at least one Difference mapping d: R ^{h × k} × R ^{h × k × l} → R ^{h × k} compares and obtains at least one difference state x _diff = d(x _test , x _ref ) (C5), ◯ the difference state x _diff according to a mapping f _score : R ^h×k → R assigns an anomaly value (AS) (C6); • Detecting the anomalies (NOK) based on the anomaly value (AS) (C7). procedure after claim 1 , wherein the reference maps, the position maps and/or the map for the anomaly value (AS) are determined by means of hyperparameter optimization. procedure after claim 1 or 2 , where • the reference state is calculated using 2 reference maps, each with a position map, with a first position map being a mean value function and a second position map being a standard deviation, or • the reference state being calculated using a reference map with a position map, with the position map being a quantile function . procedure after claim 3 , where the difference state x _diff is calculated according to x _diff = x _test - (x _mean + 3 x _std ), where x _mean and x _std are the proportions of the reference state x _ref resulting from the calculation of the mean and the standard deviation via result from the training conditions. Method according to one of the preceding claims, wherein m differential states are calculated by means of m differential images and the anomaly value is assigned to the m differential states according to a mapping f _score : R ^h×k×m → R . Method according to one of the preceding claims, in which the anomaly value (AS) is calculated as the mean value of all positive values in the difference state or states x _diff . Method according to one of the preceding claims, wherein the h ■ k positions correspond to coordinates in a 2-dimensional space. Method according to one of the preceding claims, wherein the training states and/or the state to be checked are x _test sonagrams. System (IF-Anom) for anomaly detection in industrial manufacturing processes (IF) comprehensive • a first module (IF-Anom Core), the data (Data1, Data2, Data3) of states (Z) of a process, a component and / or a production machine (PM, P) in at least one process step (PS) to be monitored industrial manufacturing process (IF) at a specific point in time or in a specific time interval, from similar process steps and/or from downstream processes, comprising data measured by sensors and/or data obtained from simulating the process, which processes the data and creates an analysis model (ANOM-Stat ) trained and maintained according to any one of the preceding claims; • a second module (IF-Anom anomaly detector), which is provided with the trained analysis model (ANOM-Stat) and which calculates an anomaly value; • a third module (IF-Anom Annotation Manager) that provides the anomaly score to a verification instance (Expert) that, ◯ if the obtained anomaly score (AS) classifies an anomalous state (Z) with high accuracy, reports the state (Z) as an anomaly without further verification; ◯ if the obtained anomaly value (AS) classifies an abnormal state (Z), determines a similarity to a state (Z) that has already been recognized as a potential anomaly and annotated, ◯ in the case of high similarity, adopts an annotation of the state (Z) previously recognized as an anomaly (NOK), ◯ otherwise annotated the state (Z) depending on a size of the obtained anomaly value (AS). ◯ annotates a condition classified as normal with a certain probability. Method for detecting anomalies in an acoustic final test of a transmission, comprising the steps • driving the transmission with a speed profile, measuring the resulting structure-borne noise using acoustic sensors and obtaining structure-borne noise time series (E1); • Transforming the structure-borne noise time series into a frequency space and obtaining sonagrams (E2); • Train an analysis model (ANOM-Stat) according to one of the Claims 1 until 8th (E3), wherein the training states include historical sonagrams of pass and fail tested transmissions and the selection of reference maps, position maps, difference maps and/or maps for an anomaly value (AS) of the analysis model (ANOM-Stat) during training data driven; • Detect an anomaly based on the anomaly score (AS) (E4). procedure after claim 10 , where • the reference figure corresponds to the calculation of the 99% percentile per position; • the difference mapping corresponds to the calculation of the difference between a state to be checked x _test and a reference state x _ref ; and/or • the mapping for the anomaly value corresponds to the calculation of the mean of all positive values in the difference state x _diff obtained. Procedure according to one of Claims 10 or 11 , wherein to carry out the method according to a system claim 9 is used.

Images