WO2010006928A1 - Procédé et dispositif de contrôle et de détermination des états d'un détecteur - Google Patents

Procédé et dispositif de contrôle et de détermination des états d'un détecteur Download PDF

Info

Publication number
WO2010006928A1
WO2010006928A1 PCT/EP2009/058300 EP2009058300W WO2010006928A1 WO 2010006928 A1 WO2010006928 A1 WO 2010006928A1 EP 2009058300 W EP2009058300 W EP 2009058300W WO 2010006928 A1 WO2010006928 A1 WO 2010006928A1
Authority
WO
WIPO (PCT)
Prior art keywords
sensor
state
classifier
data
features
Prior art date
Application number
PCT/EP2009/058300
Other languages
German (de)
English (en)
Inventor
Thomas Alber
Edin Andelic
Detlev Wittmer
Martin Freudenberger
Original Assignee
Endress+Hauser Conducta Gesellschaft Für Mess- Und Regeltechnik Mbh+Co. Kg
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Endress+Hauser Conducta Gesellschaft Für Mess- Und Regeltechnik Mbh+Co. Kg filed Critical Endress+Hauser Conducta Gesellschaft Für Mess- Und Regeltechnik Mbh+Co. Kg
Publication of WO2010006928A1 publication Critical patent/WO2010006928A1/fr

Links

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01D—MEASURING NOT SPECIALLY ADAPTED FOR A SPECIFIC VARIABLE; ARRANGEMENTS FOR MEASURING TWO OR MORE VARIABLES NOT COVERED IN A SINGLE OTHER SUBCLASS; TARIFF METERING APPARATUS; MEASURING OR TESTING NOT OTHERWISE PROVIDED FOR
    • G01D3/00—Indicating or recording apparatus with provision for the special purposes referred to in the subgroups
    • G01D3/08—Indicating or recording apparatus with provision for the special purposes referred to in the subgroups with provision for safeguarding the apparatus, e.g. against abnormal operation, against breakdown
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N27/00—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means
    • G01N27/26—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating electrochemical variables; by using electrolysis or electrophoresis
    • G01N27/416—Systems
    • G01N27/4163—Systems checking the operation of, or calibrating, the measuring apparatus
    • G01N27/4165—Systems checking the operation of, or calibrating, the measuring apparatus for pH meters

Definitions

  • the invention relates to a method and a device for checking and determining states of a sensor which, for example, supplies sensor data in a processing process.
  • a sensor which, for example, supplies sensor data in a processing process.
  • current and future states of the affected sensor which can be a pH sensor, in particular, but not exclusively, determined and made available.
  • the desired sensor data such as the pH, a temperature value or the like.
  • it is desirable to obtain as accurate information as possible about the actual instantaneous state of the sensors used in the production process, without having to remove them and separately examine them in a maintenance operation.
  • German Patent DE 10 2004 012 420 B4 discloses, for example, a monitoring device for loading sensors by influences from the measuring environment.
  • the sensors are monitored here in operation on the basis of sensor data and stored predefined value range pairs, wherein a load detection unit is provided, on the basis of which a load index for the sensors can be determined.
  • a load detection unit is provided, on the basis of which a load index for the sensors can be determined.
  • two measured variables such as the pH value and the temperature
  • the load index can be visually displayed for the sensor.
  • This known monitoring device is based on the physically measured sensor values, so that a statement about the load is possible, but the current instantaneous or future state of the sensor can not be determined here.
  • a method for checking and determining states of a sensor by means of which sensor data can be detected in processing processes or production plants and whose sensor properties together with the measured sensor data form a multi-dimensional sensor data vector V s , the sensor data vector V s for determining statements about the state the sensor itself is used and wherein the method is characterized by the steps:
  • the method according to the invention makes it possible, with these steps a) and b), to determine states in the form of a one-dimensional coordinate value and to present them clearly.
  • the determining the condition of the sensor Classification value Z s is determined on the basis of features generated in a feature generator and / or a sensor data vector V 3 with sensor data and known sensor properties.
  • the sensor data which can be detected anyway by the sensor and additional features and properties relating to the sensor are used according to the invention as an input for a state conditioner which uses a specific classification method:
  • the condition classifier according to the invention for determining current or future states of a sensor during operation or in a running processing process based on methods trained with learning methods.
  • the state classifier according to the invention may be any classifier based on trained learning methods.
  • Examples of a classifier according to the invention are artificial neural networks or a support vector machine developed in the field of statistics.
  • the basis of the determination of states is no longer the physical or heuristic modeling of actual states as in the prior art, but according to the invention the predictions and statements about the actual state of the sensor are learned data-driven, ie on the basis of actually acquired sensor data and in any case additionally created by the sensor detectable characteristics, properties and data with learning methods.
  • sensor data can be understood to be any data that can be sensed. That is, the invention is not limited to the application in the state detection of sensors, but any System may relate to actuators or devices in which sensory sizes are determined.
  • condition classifier trained on the basis of learning methods is to be understood as meaning a module or a program in which a functional is learned beforehand or during ongoing operation, which is then filled with acquired sensor data or further features in the form of data vectors.
  • the state determination can also be learned initially independently of the later operation. With a pre-training before the later operation in the sensor itself or in the sensor-related transmitter initially a sufficient basis is created to allow a very accurate statement about the current or the future state of the sensor.
  • the invention thus provides a more meaningful and accurate state classification and state prediction of sensors on the basis of learning methods; in contrast to the purely physically based methods from the past, which are based on periodically or permanently acquired data, a data-driven learned in the inventive method It also has the advantage that it is not time-critical and can, for example, calculate a meaningful classification result at any given time.
  • the method of the present invention provides a one-dimensional classification value Zs representative of a condition of the sensor.
  • a classification value includes a value which makes a direct and clearly visualizable statement about the current or future state: in one possible embodiment, the one-dimensional classification value Zs can be represented in the form of a traffic light, ie red for malfunction , yellow for maintenance and green for correct operation and proper functioning of the sensor.
  • the features for the feature generator in the sensor itself are extracted and / or determined.
  • the features which are required for the state classification according to the invention and which In addition to the sensor data serve as input to the classifier, are extracted or determined in this case in the sensor itself.
  • the feature extraction is performed in the sensor itself, it is possible to sample at a faster sampling rate and at a faster rate than would otherwise be possible. In this way, occurring noise effects in the data acquisition can be better determined, which contribute to increasing the informational value of determined state statements.
  • an m-dimensional feature vector V M is generated as an output, which is supplied as an input to the state of the invention according to the invention.
  • the m-dimensional feature vector V M includes features that can be determined directly or indirectly via the sensor, such. B. the sensor data recorded as raw data and known properties of the sensor. From the sensor data, a multi-dimensional feature vector is generated in the feature generator, which is supplied for further processing and actual classification as input to a trained state conditioner according to the invention.
  • the feature vector V M may have a same dimension or a higher dimension compared to the sensor data vector Vs.
  • the features and feature values contained in the feature vector V M are matched to a proper check and determination of the condition of the sensor.
  • As feature quantities in the feature vector V M derived quantities based on actuators, operators or the like can also be used.
  • operators are used to obtain the state relevant to the sensor srelevanten features or feature values based on sensor data and / or sensor properties.
  • Operators in the present context are understood as meaning devices or program modules by means of which further properties and features can be determined that are related to the direct sensor data and the directly detectable with the sensors sensor properties can be used in a useful manner for the determination of state. Examples of so-called operators are the variance of a signal or other functionalities, for example, for the formation of equivalences, so-called complex features or simple process or event counter, which can be determined from the measured sensor data. Examples of operators are also given in the later described embodiment of the invention.
  • a data-driven learned functional for performing the state classification and / or state prediction with respect to the sensor is used in the condition classifier.
  • the functional according to the invention is a data-driven learned functional whose training is based on a learning method.
  • the functional according to the invention in the state classifier is thus not modeled physically or heuristically as in the prior art, but is based on a training on the basis of actual sensor data and features that are generated with the steps according to the invention.
  • the functional learned in advance in the state classifier or in the operation according to the invention is then filled in the actual operation and the actual making the state classification with a feature vector V M.
  • training according to the invention can first be carried out independently of the later operation in the sensor or in the transmitter.
  • the actual verification and determination of states then carried out by the method according to the invention is realized with high degree of accuracy.
  • a functional once learned in the condition classifier which is externally performed, for example, by a training on a computer, can be advantageously used in later operation for precise status statement with respect to the sensor.
  • the training of the condition classifier for "learning" the function can, for example, also be carried out on the basis of expert systems or an expert himself Saved data from known states of sensors or typical state changes are also suitable for training the state classifier according to the invention.
  • a history of features or their extrapolation into the future is used as an input for the condition classifier for determining a future state of the sensor.
  • the history of features is understood to be the past development of feature values, which according to the invention is preferably stored for later use.
  • the extrapolation of a history of characteristics or feature prediction can be done on the basis of common mathematical or statistical calculation methods. Thus, it is possible according to the invention to make an accurate future statement about states of the sensor at certain times.
  • the senor is operated with a higher compared to the normal operation of the sensor sample rate. Due to the higher sample rate or sampling rate, the sensor can provide more state-relevant data and features than would be the case at normal Sampie rate during operation without the state classification according to the invention. With the inventively inflated sample rate, additional data can be provided with one and the same sensor without additional equipment expenditure, which data can be used advantageously for the process diagnosis and in particular the diagnosis of the condition of the sensor itself. Furthermore, this aspect of the invention makes it possible to differentiate between a higher sample rate between a sensor dynamic and a process dynamic, and by corresponding analyzes of the data, changes in the two dynamics can be detected quickly and effectively in good time.
  • the condition classifier is trained with training feature vectors and / or training features which reflect known state situations of the sensor.
  • the training feature vectors and / or Trainzansmerkmale can be stored and provided, for example in the form of an expert system.
  • an external as well as an internal storage of such training data - based on the sensor device or its control - is equally possible.
  • methods such as a support vector machine, abbreviated SVM, a neural network, a discriminant analysis or a Gaussian process are used in the trained condition classifier as the learning method.
  • SVM support vector machine
  • neural network a discriminant analysis or a Gaussian process
  • Gaussian process a feature-based process
  • the invention is not based on a physical or heuristic model, but explicitly uses a data-driven learning model, which is also known by the term "training of a learning method" in fields other than that of the present invention used in the context of a condition classifier for sensors, which thus advantageously a higher classification accuracy and better predictions and predictions about current and future states of the sensors are possible.
  • the data-driven learned functional of the condition classifier and / or the evaluation for condition classification are stored or performed in a transmitter of the sensor.
  • the execution of the function and the storage is in the transmitter usually enough space available, at least more than it is in the sensor Fail.
  • Sufficient memory available in the transmitter is an important consideration when performing the necessary calculations and classification operations based on a large amount of data.
  • the evaluation of the function of the present invention is not time critical. It can be executed at arbitrary times, and it is not necessary to immediately calculate a classification result at each sampling instant of the sensor. For example, according to the method of the invention it is sufficient to consider the classification every few minutes and to view the result of the condition classifier.
  • the transmitter is therefore according to one embodiment of the present invention, the preferred location for storing and performing the method steps,
  • an extrapolation of the sensor state based on a state history of the sensor is performed to predict future states of the sensor.
  • the history of Merkmai characteristics with respect to the state of the sensor can be swept over a longer period, for example, in a memory and used for extrapolation or prediction of future conditions.
  • Statements about the future state of the sensor can also be determined in particular from statistical methods and methods for detecting trends in the characteristic values of the sensor.
  • a device for checking and determining states of a sensor having the features according to claim 13 is also proposed.
  • the inventive A state determination device comprises at least one feature generator and a state identifier, wherein the sensor is provided with a transmitter and wherein the state identifier is a classifier trained on the basis of learning methods, which is adapted to generate a one-dimensional classification value Z s from features generated in the feature generator and / or generate other data, the one-dimensional classification value being representative of the current or future state of the sensor.
  • the state check device according to the invention enables a very accurate determination of current and future states of the sensor, whereby a one-dimensional classification value is delivered as a result.
  • the one-dimensional classification value can easily be visualized, for example in the form of a status display designed as a traffic light.
  • the device according to the invention is simple in construction and does not require any additional equipment in comparison with previous devices of this type. With one and the same sensor and on the basis of existing data, according to the invention only a larger recordable data base than before is used in order to obtain a meaningful statement about the states of the sensor.
  • the condition as well as the feature generator can be accommodated, for example, as program modules within an existing control and control unit.
  • additional memory modules and processing modules are to be provided on the sensor, on the transmitter or as separate units, but communicating with each other via data lines.
  • the invention has the advantage that an expansion and periodic maintenance of the sensor are no longer required as often as before. Faulty states can be detected early, and even in operation can quickly identify situations based on feature characteristics and trends, such as a soon required maintenance work or a complete replacement of the affected sensor at a given time.
  • a memory unit for storing training feature states of the sensor and / or a learnedmetsais is provided.
  • the memory unit advantageously contains Merkmaisaus Weggungen and corresponding associated states, so that on the basis of this database similar to an expert system, the device or its state conditioner can be trained on the basis of existing data.
  • the memory unit also allows storage of further acquired features and sensor data and sensor properties during operation, so that increasingly over time the underlying data base for state evaluation is optimized and thus the reliability of the classification statements with respect to sensor states is improved more and more.
  • a memory for storing feature vectors V M and / or further sensor data is provided, which is spatially separable from the sensor and the transmitter.
  • a training for learning the required for the classification function based on learning methods can be obtained offline and separately from the sensor itself.
  • a function for the state identifier with the memory can be trained in advance before using the sensor. Subsequently, the functional can be taken over in the sensor, for example in a control and regulating device and stored there, for example in the form of a program module.
  • an extrapolator is provided, which is adapted to determine a state prediction with respect to the sensor on the basis of a stored history of features or statistical methods.
  • the device according to the invention allows not only a determination of the current state, but also meaningful state classifications for certain times in the future, which are each freely selectable by the operator. An early one Detecting a necessary replacement of the sensor is thus possible.
  • the senor is a pH glass electrode for measuring pH values.
  • the condition classifier comprises as a means of classification a support vector machine, SVM 1 a neural network, a discriminant analysis or a Gaussian process as a learning method in the condition classifier, which is incorporated in the device, for example in the form of program modules are.
  • Fig. 1 is a schematic block diagram for illustrating the construction of a first embodiment of an apparatus for condition classification of a sensor according to the invention.
  • FIG. 2 is a schematic block diagram of a second embodiment of a device for condition classification and state prediction of sensors with an extrapolator for predicting future states, according to the present invention
  • Fig. 1 shows schematically the structure of a first embodiment of a device according to the invention for checking and determining the states of a sensor.
  • a sensor 3 is connected to a trained state classifier 2 according to the invention and to a feature generator 1.
  • the sensor 3, which may be, for example, but not exclusively, a pH sensor, has a measuring transducer 4, which is also connected to the condition classifier 2 communicates.
  • a memory 5 is provided in which sensor data and further data as well as a trained functional for the condition classifier 2 can be stored.
  • the device according to the invention is used to detect and check the state of the sensor 3, wherein both the current state and future states of the sensor 3 can be determined with the device according to the invention.
  • a specific trained condition classifier 2 is provided, which was trained on the basis of a learning method. On the basis of sensor data which can be determined with the sensor and further known properties of the sensor 3, an evaluation is made in the condition classifier 2 in such a way that a one-dimensional value which is representative of the sensor state can be determined.
  • state value Zs This is shown in the figures with the state value Zs and a display in the form of a simple traffic light.
  • the state classification and prediction of future states according to the method according to the invention takes place in the trained state classifier 2 on the basis of learning methods, in the following manner: With the sensor 3, in addition to the normal sensor data, further data can be detected, which are not used in normal operation.
  • the sensor data and this further data that can be detected by the sensor 3 are input into a feature generator 1 together with known properties of the sensor as an m-dimensional sensor data vector V s .
  • the sensor data vector V s thus consists of the measurable raw data, which are available through the sensor 3, as well as other known properties of the sensor 3.
  • the sensor data vector Vs consists of features obtainable by operators, which in the feature generator specifically with regard to the state evaluation and be generated as needed.
  • the feature vector V M is supplied as an input to the condition classifier 2.
  • the condition classifier 2 is particularly adapted for the state determination and state prediction with respect to the sensor 3 and is based on a so-called learning method.
  • the condition classifier is preferably trained beforehand, ie before the actual operation, with known state sensor data, an expert system or an expert, and equipped with a functional, via which a one-dimensional sensor state value can be determined.
  • the condition classifier 2 consists for example of an artificial neural network, a support vector machine (SVM) or a method based on discriminant analysis or Gaussian processes.
  • SVM support vector machine
  • One of these learning methods is used in the condition classifier 2 to provide a meaningful classification regarding the condition of the sensor 3.
  • this results in a one-dimensional sensor state Z s , which can be visually displayed in a display device, in the simplest case for example by a traffic light with red, green and yellow lighting device, as shown schematically in FIGS. 1 and 2 , Depending on the detected condition, the traffic light is switched to one of the three values.
  • Other means of representation with one-dimensional display are also conceivable.
  • An exemplary embodiment of a state classifier 2 trained on the basis of learning methods for determining sensor states can be realized in the following manner.
  • a so-called functional is generated on the basis of sensor data and known state feature expressions, which is used in the subsequent operation of the state identifier 2 to make predictions about current and future states of the sensor at certain times.
  • a sensor state is thus no longer modeled physically or heuristically.
  • the sensor state is rather with the state classifier according to the invention learns data driven, with a so-called learning method. Under learning method in the present context, the training of aginaais for the sensor on the basis of data from the sensor, ie data-driven understood.
  • the method of the Support Vector Machine is preferably used in the context of the present invention, as described, for example, Schölkopf / Smola: Learning with Kernels: Support Vector Machines, Regularizatlon, Optimization and Beyond (Adaptive Computation and Machine Learning), MIT Press, Cambridge, 2002.
  • the Support Vector Machine is used to learn a functionally data-driven based on data that can be acquired with the sensor 3 and other features.
  • This functional is entered into the condition classifier 2 after the training with the aid of the learning method and used in the further operation of a production process, for example, to permanently or at certain times evaluate the condition of the sensor 3. This can with the device according to the invention sowoh!
  • first feature vectors and two additional scalar parameters are calculated as support vectors for the support vector machine, by which the function for condition classifier 2 is then completely defined.
  • only labeled data of the sensor is used, ie the detectable actual sensor data as well as further data and characteristics which can be determined with the sensor, such as the properties known via the sensor.
  • the functional once learned via the Support Vector Machine is used to determine the state of the sensor.
  • a feature extraction for the feature generator 1 can be carried out according to an advantageous embodiment of the invention in the sensor 3 itself, since it can be scanned faster and thus also noise effects can be better detected.
  • measured values with an excessive sample rate are detected at the sensor 3 for this purpose.
  • an excessive sample rate or sampling rate at the sensor 3 can In particular, short-term trends and a noise characteristic, for example due to influences from the measurement environment, can be recognized and properly assessed with regard to status statements.
  • condition classifier 2 the actual classification takes place, ie.
  • a feature vector V M generated with the feature generator 1 is inserted into the above-mentioned learned functional of the state classifier 2, whereby the affiliation of this feature vector V M to a particular class in the form of a scalar is displayed in a display device becomes.
  • the result is a one-dimensional value for the state, the sensor state Z s , which in turn can be represented in the simplest form, for example as a traffic light.
  • the execution of the function of the condition classifier 2, in the case of a support vector machine, ie the storage and calculation of all support vectors and the corresponding scalar parameters, is advantageously carried out in the transmitter 4 of the sensor 3.
  • the evaluation is also advantageously carried out in the transmitter 4.
  • the inventive evaluation of the function and determination of state classification values is not time-critical, and it is not necessary, for example, to calculate a classification result for each sampling time of the sensor 3. It is sufficient according to the present invention, for example, every few minutes to determine and display a Klassäfikationswert Z s . This is best done in the transmitter 4 of the sensor 3.
  • FIG. 2 shows a second exemplary embodiment of a device according to the invention for determining the state of sensors on the basis of a schematic block diagram.
  • a state predictor ie a prediction of future conditions of the sensor 3 at certain future times. This can be a necessary replacement of Sensor elements or required maintenance can be predicted.
  • the core of this device is also a condition classifier 2 trained on the basis of so-called learning methods. The statements made above with regard to the condition classifier 2 also apply to the condition classifier 2 according to the exemplary embodiment of FIG. 2 too.
  • the two embodiments of FIGS. 1 and 2 can also be realized together and combined in one and the same device. They are described separately in the present context merely for reasons of presentation and to facilitate understanding.
  • a history of features is additionally stored here. This is shown in FIG. 2 with the history memory module! 7 illustrated.
  • a feature vector V M is first generated in a feature generator starting from a sensor data vector V 3 .
  • a stored history 7 of features relating to the sensor 3 is used, which are used together as an input into an extra-poiator 6.
  • an extrapolation of the features to the future is performed on the basis of the history of features 7 and of the feature vector V M.
  • the state classifier 2 determines a one-dimensional state value Z s for the state of the sensor 3, but here for a future state at a specified future point in time. This is then again in the form of a simple means of representation, such. As a traffic light, visualized, as is schematically indicated in FIG. 2.
  • future sensor states and their times can also be predicted in advance. The failure probability of the sensor is thus significantly reduced compared to previous methods for state evaluation.
  • membership probabilities for possible sensor states are also stored and used for a state prediction in the condition classifier 2.
  • the method according to the invention is carried out in the form of an algorithm with the aid of a trained function, which then runs in the state classifier 2 during later operation.
  • a monitoring, recognition and assessment of process changes with respect to the sensor 3 are advantageously carried out not in the sensor itself, but in the transmitter 4 of the sensor 3.
  • the state determination according to the invention and the state prediction with respect to the sensor can be combined with a control or regulation device within a production process or also be integrated into it.
  • a pH glass electrode is used in production processes for measuring pH. It consists of two half cells, each filled with an internal buffer.
  • One of the two half-cells is connected via a pH-sensitive glass membrane, wherein the other of the half-times is connected via a so-called diaphragm with the process medium.
  • the difference between the potentials of the two half-cells is a measure of the pH of the process medium within a production process in which the sensor 3 is used.
  • other sensor properties must also be considered for the calculation, which are given in the form of zero point and slope, for example are. Zero point and slope can not be measured in situ, ie in situ, in the process until today, and therefore must be cumbersomely determined by adjusting the sensor 3 with the aid of buffer solutions having a defined pH value.
  • a simple operator that can advantageously be used for the method according to the invention is, for example, the variance of the measuring signal of the sensor 3. This is illustrated mathematically below using an example, where X 0 is the data vector, X M is a possible feature vector and X z is a state vector with four here possible manifestations of the state classification.
  • More complex operators for generating additional features for the condition classification of the sensor 3 according to the invention are, for example, functionalities for the formation of equivalents, so-called complex features such as Merkmai BA in the feature vector X M in the above example. Often, simple features such. As event or event counter, helpful for a good classification performance for condition classification of sensors in the present invention.
  • the feature vector X M shown here contains not only elements from the data vector X D and the sensor data present there as well as characteristics of the sensor 3, such.
  • B the potential of the reference half cell ⁇ R , and the shifts of zero and slope of the last two adjustments ⁇ Np, ASt, the variance of the time series of some variables var (x) and also a functional in the form of a load equivalent BA, which statements about the load state of Sensor 3 contains.
  • the object of the feature generator 1 is, by using various operators, the feature vector V M or in the Beispie! X M from the data vector X 0 or to calculate its history.
  • the state classifier 2 now has the task of mapping the current feature vector X M , which represents the current sensor state better than just the data vector X D , to a discrete state equation as a result of the method according to the invention.
  • the possible state classes of the state vector X z are in principle freely selectable, and it is irrelevant whether the selected name reflects an actual state or a future state or stands for a maintenance measure to be carried out.
  • meaningful condition classes can be, for example, the "ok”, “adjust” the required maintenance measures, "clean / regenerate” or the measure "exchange” in terms of worn sensor, as described above
  • Example of a state vector X z is the case.
  • Decisive for the Performance of a state classifier 2 according to the invention is the choice of methodology according to which the mapping of the feature vector X M in the state vector occurs, as well as the choice of the classifier 2 itself compared to model-based approaches or knowledge-based systems such as expert systems have learning methods according to the present invention clear advantages.
  • the state classifier 2 according to the invention can consist of, for example, artificial neural networks or a support vector machine SVM, which in the case of more complex sensors leads to better statements about the current and future states.
  • condition classifier 2 For the training of condition classifier 2 in advance, ie before commissioning of the production process, based on learning methods, data vectors are first recorded and classified by hand by an expert. In this case, one also speaks of a labein the data, so that expert-classified, labeled data arise. The condition classifier 2 is then trained to learn the classification of the expert. However, such a labein the data by hand is only possible in a limited way. On the one hand, it should only be labeled if the current state of the sensor 3 can really be detected by the expert with a high degree of certainty, so that the later classification mode is not falsified.
  • condition classifiers with a high generalization capability.
  • the more reliably a condition classifier classifies 2 unknown data-that is, data that was not used for training-the higher its generalization capability.
  • the object of the present invention is therefore to propose a condition classifier which guarantees this high reliability with as little training data as possible in order to obtain the best possible results even with a smaller number of data determining the current or future state of a sensor.
  • Support Vector Machine As an example of such a state classifier according to the present invention may be mentioned, inter alia, a methodology called Support Vector Machine.
  • Support Vector Machines SVM for short, have been described in fields other than the present in other application contexts in the literature. Support vector machines have emerged from the findings of statistical learning theory and offer significant advantages in terms of their ability to generalize in the context of the present invention.
  • the classic basis for the training of a support vector machine is a set of training objects, for each of which it is known to which state classes with respect to a specific sensor 3 they belong. This is commonly referred to as a supervised learning. Each object is represented by a vector in a vector space.
  • the task of the support vector machine of state classifier 2 is to find in this vector space a multidimensional hyperplane that can act as a breaker and subdivides the training objects into two different classes, with a problem consisting of more than two classes easily passing through a fusion of several two-class problems is solvable.
  • the distance to the hyperplane of those vectors closest to the hyperplane is thereby maximized.
  • the existence of this broad, empty margin, mathematically, implies that even objects that do not exactly correspond to the training objects for this sensor 3 are reliably classified.
  • This idea can be formulated mathematically as a quadratic optimization problem in which a total of two scalar parameters must be optimized during the training of the condition classifier 2.
  • the solution of this optimization problem shows that it is not necessary for the hyperplane sought to observe all training vectors. Training vectors that are farther from the hyperplane and, as it were, "hidden" behind a front of other vectors do not affect the location and position of the separation plane, so the hyperplane is only of the vectors closest to it Depending - and only these are needed to describe the level in the method according to the invention mathematically exact. These closest vectors are also classified as function vectors according to their function. support vectors.
  • the support vectors are always a subset of the training vectors, their position is automatically determined by the solution of the mathematical optimization problem described above and therefore does not need to be trained.
  • This approach also has the advantage that future series of measurements of the sensor 3 can be better planned by making particularly many measurements with the sensor 3, especially near the respective support vectors.
  • support vector machines in the present invention can be trained even if only a part of the training data is forked.
  • the function of the unlabeled data is to check, so to speak, the plausibility of the result from a condition classification.
  • This learning scenario is also referred to elsewhere in the literature as semi-supervised tearning.
  • the kiasshuis intricate with the state classifier based on a support vector machine according to the invention can far exceed that of an expert, especially in the case of difficult to classify data sets of more complex sensor systems.
  • Such a trained state classifier 2 according to the invention can advantageously be used in addition to the assessment of the current state for a state classification in the future, ie for a prediction of future states at specific times.
  • continuations can be estimated from statistical properties known to those skilled in the art for particular types of sensors.
  • the correlation of the individual features with each other or the relevance for their prediction in the future can be read from the trained condition classifier 2 according to the present invention. Accordingly, it is only necessary to estimate the relevant characteristics.
  • a feature vector thus determined for any desired time in the future is now in turn supplied to the condition classifier 2 according to the present invention, and a state classification is thus obtained for a specific future time in the form of a future state value Zs in one-dimensional form.
  • a state classifier 2 operates internally numerically and first calculates for a feature vector a measure of membership in each of the output lanes possible in the current case. The numerically largest membership then determines the condition class to which the feature vector is assigned or which is output at the output.
  • One possibility according to the invention is now to store the numerical affiliation with the possible state classes over time in a memory module. On the basis of this stored membership history, the expected future affiliations can then be estimated by means of regression or extrapolation. These estimated class memberships may then be supplied to decision makers or operators of the inventive apparatus at the output of the condition classifier, and the associated condition class for the selected time in the form of a future state value Z s .
  • Both methods can thus be used to determine future states or class membership for different points in time in the future in a simple manner with relatively little computational effort. For example, for the evaluation of sensors of interest are those times at which a transition from one class to another will take place.
  • This state prediction according to the invention with the point in time and the state associated therewith thus makes an easily understandable statement, for example on the basis of a traffic-type indication as to when which maintenance measures on the sensor 3 will be required in the future.
  • the invention offers increased process reliability and reduces sensor-related failures in production processes.

Abstract

L'invention concerne un dispositif et un procédé destinés à la classification des états des détecteurs. Le procédé selon l'invention sert au contrôle et à la détermination des états d'un détecteur, au moyen duquel des données de détecteur puissent être détectées, en tant que données brutes dans un processus de traitement, et dont les propriétés de détecteur forment, conjointement avec les données de détecteur mesurées, un vecteur de données de détecteur multidimensionnel Vs , le vecteur de données de détecteur Vs étant utilisé pour la détermination d'information sur l'état du détecteur. Le procédé selon l'invention est caractérisé en ce qu'il comprend les étapes suivantes : a) génération de caractéristiques dans un générateur de caractéristiques (1) en se basant sur au moins le vecteur de données de détecteur Vs; b) utilisation d'un classifieur d'états (2) entraîné sur la base de méthodes d'apprentissage, pour la détermination d'une valeur de classification monodimensionnelle Zs, qui est représentative de l'état actuel ou futur du détecteur (3).
PCT/EP2009/058300 2008-07-14 2009-07-02 Procédé et dispositif de contrôle et de détermination des états d'un détecteur WO2010006928A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
DE200810032885 DE102008032885A1 (de) 2008-07-14 2008-07-14 Verfahren und Vorrichtung zur Überprüfung und Feststellung von Zuständen eines Sensors
DE102008032885.5 2008-07-14

Publications (1)

Publication Number Publication Date
WO2010006928A1 true WO2010006928A1 (fr) 2010-01-21

Family

ID=41306911

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/EP2009/058300 WO2010006928A1 (fr) 2008-07-14 2009-07-02 Procédé et dispositif de contrôle et de détermination des états d'un détecteur

Country Status (2)

Country Link
DE (1) DE102008032885A1 (fr)
WO (1) WO2010006928A1 (fr)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3457106A1 (fr) * 2017-09-13 2019-03-20 Rolls-Royce Deutschland Ltd & Co KG Procédé et dispositif de surveillance de la cinématique d'un engrenage planétaire épicycloïdal

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE102013114393A1 (de) * 2013-12-18 2015-06-18 Endress + Hauser Conducta Gesellschaft für Mess- und Regeltechnik mbH + Co. KG Verfahren zur Bestimmung einer Wartungsempfehlung für einen Sensor
DE102019107625A1 (de) * 2018-12-20 2020-06-25 Endress+Hauser Conducta Gmbh+Co. Kg Verfahren zur In-Prozess-Justage eines potentiometrischen Sensors einer Messanordnung
DE102019219300A1 (de) * 2019-12-11 2021-07-01 Robert Bosch Gmbh Ermitteln relevanter Sensoren für die Zustandsüberwachung von Geräten und Systemen
DE102019219301A1 (de) * 2019-12-11 2021-06-17 Robert Bosch Gmbh Reduzierung des Datenvolumens bei der Zustandsüberwachung von Geräten und Systemen
DE102022205534A1 (de) 2022-05-31 2023-11-30 Kuka Deutschland Gmbh Überwachen einer mehrachsigen Maschine mittels interpretierbarer Zeitreihenklassifikation

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE4434197A1 (de) * 1993-10-04 1995-04-06 Ford Werke Ag Diagnosetechnik für den Betrieb eines Abgassauerstoffsensors
EP1630634A2 (fr) * 2004-08-25 2006-03-01 Siemens Corporate Research, Inc. Procédé et dispositif de détection des conditions "hors limites" dans des systèmes de génération d'énergie
US20070288409A1 (en) * 2005-05-31 2007-12-13 Honeywell International, Inc. Nonlinear neural network fault detection system and method
DE102006030895A1 (de) * 2006-07-01 2008-01-03 Endress + Hauser Conducta Gesellschaft für Mess- und Regeltechnik mbH + Co. KG Messanordnung mit Zustandsüberwachung und Verfahren zur Zustandsüberwachung einer Messanordnung

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE4100501A1 (de) * 1991-01-10 1992-07-16 Bodenseewerk Geraetetech Verfahren und einrichtung zum erkennen von fehlern an sensoren fuer zustandsgroessen
WO2000010059A1 (fr) * 1998-08-17 2000-02-24 Aspen Technology, Inc. Procede et dispositif de validation de capteur
DE102004012420B4 (de) 2004-03-13 2007-03-01 Knick Elektronische Messgeräte GmbH & Co. KG Überwachungsvorrichtung für die Belastung von Messsonden durch Einflüsse aus der Messumgebung

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE4434197A1 (de) * 1993-10-04 1995-04-06 Ford Werke Ag Diagnosetechnik für den Betrieb eines Abgassauerstoffsensors
EP1630634A2 (fr) * 2004-08-25 2006-03-01 Siemens Corporate Research, Inc. Procédé et dispositif de détection des conditions "hors limites" dans des systèmes de génération d'énergie
US20070288409A1 (en) * 2005-05-31 2007-12-13 Honeywell International, Inc. Nonlinear neural network fault detection system and method
DE102006030895A1 (de) * 2006-07-01 2008-01-03 Endress + Hauser Conducta Gesellschaft für Mess- und Regeltechnik mbH + Co. KG Messanordnung mit Zustandsüberwachung und Verfahren zur Zustandsüberwachung einer Messanordnung

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3457106A1 (fr) * 2017-09-13 2019-03-20 Rolls-Royce Deutschland Ltd & Co KG Procédé et dispositif de surveillance de la cinématique d'un engrenage planétaire épicycloïdal

Also Published As

Publication number Publication date
DE102008032885A1 (de) 2010-01-21

Similar Documents

Publication Publication Date Title
DE102017011290B4 (de) Vorrichtung für maschinelles Lernen, CNC-Vorrichtung und maschinelles Lernverfahren zum Erkennen eines Anzeichens eines Auftretens von Rattern im Werkzeug einer Werkzeugmaschine
DE102017108169B4 (de) Produktionssystem, das einen Bestimmungswert einer Variablen in Bezug auf eine Produktabweichung festlegt
DE102017000536B4 (de) Zellsteuereinheit zum Feststellen einer Ursache einer Anomalie bei einer Fertigungsmaschine
EP1543394B1 (fr) Procede et dispositif pour surveiller une installation technique comprenant plusieurs systemes, notamment une centrale electrique
DE102019003382A1 (de) Anomaliedetektor
DE102019128177A1 (de) Vorrichtung und verfahren zur zustandsbestimmung
DE102018108779A1 (de) Schätzvorrichtung eines maschinellen Bearbeitungsfehlerfaktors
DE102018006024A1 (de) Controller und maschinelle Lernvorrichtung
DE102006048430A1 (de) Verfahren zur Wartungsvorhersage einer Maschine
DE60211520T2 (de) Anordnung zur Steuerung des Betriebs eines physikalischen Systems wie zum Beispiel einer Brennstoffzelle in elektrischen Fahrzeugen
DE102017107284A1 (de) Verfahren und steuergerät zum überwachen eines bordnetzes eines fahrzeugs
DE102016011523A1 (de) Motorantriebsvorrichtung mit vorbeugender Wartungsfunktion eines Gebläsemotors
WO2010006928A1 (fr) Procédé et dispositif de contrôle et de détermination des états d'un détecteur
DE102017000287A1 (de) Zellensteuerung und produktionssystem zum verwalten der arbeitssituation einer vielzahl von fertigungsmaschinen in einer fertigungszelle
DE102017003427A1 (de) Produktionssystem zur Durchführung eines Produktionsplans
DE102020100668B4 (de) Charakterisierung von wiederaufladbaren Batterien mit Maschinen-gelernten Algorithmen
EP3282399A1 (fr) Procede de reconnaissance ameliore d'anomalies de processus d'une installation technique et systeme de diagnostic correspondant
DE102019001044A1 (de) Steuervorrichtung und maschinelle lernvorrichtung
DE102018126434B4 (de) Testvorrichtung und Maschinenlernvorrichtung
DE102019001760A1 (de) Informationsverarbeitungsgerät, maschinelle lernvorrichtungund system
EP3876061A1 (fr) Procédé de validation et de sélection de modèles basés sur l'apprentissage automatique destiné à la surveillance d'état d'une machine
DE102019104922A1 (de) Kollisionspositionsschätzvorrichtung und maschinenlernvorrichtung
DE102008045840B4 (de) Verfahren zum Betreiben einer Prozess-Messstelle
DE102022202984B3 (de) Verfahren zum Trainieren eines neuronalen Netzwerks zum Ermitteln eines Merkmalsvektors, der den Abnutzungszustand einer Vorrichtung repräsentiert
DE102005038889A1 (de) Verfahren zur Prüfung eines Fertigungsgutes

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 09780082

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 09780082

Country of ref document: EP

Kind code of ref document: A1