WO2006089330A1

WO2006089330A1 - Methods and arrangement for identifying the deviation from determined values

Info

Publication number: WO2006089330A1
Application number: PCT/AT2006/000080
Authority: WO
Inventors: Brian Sallans; Gerhard Russ
Original assignee: Arc Seibersdorf Research Gmbh
Priority date: 2005-02-24
Filing date: 2006-02-23
Publication date: 2006-08-31
Also published as: AT502241A1; AT502241B1

Abstract

The invention relates to methods and a device for monitoring the deviation of determined values from setpoint values, said values being determined by means of transmitters of units that are monitored therewith and being output in the form of output values. According to the invention, a probability value for the next output value of a transmitter is assigned to each transmitter with the aid of probability models and/or the output values of at least two different transmitters are correlated with each other by establishing a probability value for the interdependence of said output values by means of probability models, and it is determined for the output values of the transmitters whether the probability value of said output value lies outside the assigned probability range or deviates from the assigned probability value.

Description

Method and device for determining the

Deviation from determined values

The invention relates to methods according to the preamble of claim 1.

Furthermore, the invention relates to an arrangement according to the preamble of claim 8.

The primary objective of the invention is the automatic detection and reporting of

Errors or deviations from nominal values or normal values of actual or output values in a sensor / actuator network which monitors functional units, in particular with regard to activity, function, behavior, appearance, etc.

As technology becomes more prevalent in everyday life, sensor networks are becoming ever more important: buildings, vehicles, airplanes, industrial equipment, and so on, can boast a multitude of sensor-monitored units networked together. Along with this development, the demand for processing and evaluating the collected information from the networks increases. Current systems require a great deal of work to process the collected data, as rules about when which values and states in sensors, devices, actuators, etc. are within the normal range must be manually created.

EP 1 418 476 shows a first step for the automated processing of information from sensor networks. EP 1 418 476 presents an adaptive control system for a vehicle which learns and interprets information about the vehicle, the environment and the user with the aid of a learning unit in order to influence the vehicle or the environment with the findings from this interpretation. The downside of this system is that it uses "supervised ^• learing." That is, the

System does not learn on its own, but it needs a teacher (supervisor) who can do it

Learning behavior is constantly monitored and controlled. Thus, the benefits of learning

Systems reduced to the already described manual configuration of the system. WO 98/56182 describes a system for the automated evaluation of camera images. Video data is captured, time-stamped, cached, and then statistically compared with data from other records that meet certain criteria. The subject of WO 98/56182 is designed for the processing of video images - methods such as grayscale analysis and Fourier transformations are used for image processing and are elementary components of this system. New data is compared to other records, but not to a model. New data must be saved in order to be able to compare them; that is, the data is captured, stored, and compared in real time with a certain number of other stored data. The system will be included System startup automatically initialized with an existing image set; after that it will be corrected manually by the user in case of deviations / changes.

The invention relates in particular to a (e) control method and

Arrangement for automation systems in buildings or installations, in particular for the automatic detection and reporting of errors or deviations from

Normal values in a sensor network, and the message of exceptional

Situations that can be detected by the sensors or by groupings of sensors. The method according to the invention should be able to bring these messages to the attention of the user or to make them available to another system (facility management system, log system, etc.).

The present invention sets itself above all the goal of being able to ascertain, in a simple and rapid and as exact as possible manner, the deviation of received measured values or output values of encoders or sensors from normal values, which measured values are output by encoders or sensors, the functional units monitor in the most general way. In question are all types of

Industrial plants, buildings, machinery, vehicles etc.

This object is achieved in a method of the type mentioned by the features cited in the characterizing part of claim 1. In an approach according to claim 1, the output values of individual sensors can be examined and assessed, which may also come from one and the same unit. However, alternatively or at the same time, a plurality of the output values obtained by the units can be correlated and assessed, resulting in a general or a possibility of assessment extending over all monitored units. The same applies to the arrangement according to claim 9.

According to the invention, the prior knowledge required in known diagnosis and warning systems about the structure of the automation system, which values belong to which sensor, actuator, control units, which correlations between these values, etc., are necessary for the creation of suitable

Control mechanisms and for entering parameters (value ranges, thresholds, etc.) avoided. The procedure according to the invention makes it possible to read value ranges, parameters and relationships from the collected information without prior knowledge, to process them and to bring them to the attention of the user.

The determination of the deviation is made online, i. without storing the values and subsequent comparison. Furthermore, not only readings of a single

Unit examines or assesses, but it can also be measured values of different entities that interact in a complex way are studied and assessed.

According to the invention, a high workload on the one hand in the installation and adjustment of the system for the manual creation of dependencies, links, etc., and on the other hand when changes in the field of application (structural changes, new users, etc.) due to manual adaptation work is avoided. The arrangement according to the invention is able to adapt to the respective area of use without any human intervention. Furthermore, it is because the Anpassungsbzw. Learning mode remains active (unless it is desired by the user to deactivate) to adapt to changes that occur over time.

For the diagnosis and for the announcement of events, which do not correspond to the normal case, it is not necessary, appropriate default settings (rules,

Shortcuts, parameters, etc.). According to the invention, messages about abnormal processes or values are generated without these being defined in advance.

According to the invention, accumulated values are processed without previous intermediate storage. New values are not compared to a stored history of values. Instead, new values are compared to a model that is a "summary" of that history, and because of this simplified comparison, and given that this comparison accounts for a large portion of the total processing time, performance can be significantly increased based on real data from the

Automation system generated or learned. Provides e.g. a sensor noisy data, so this noisy data are integrated into this model. In contrast to conventional systems which expect correct values but are then confronted with noisy data, according to the invention, these noisy data are recognized and considered normal for this sensor. Abnormal values are still recognized because they are outside the normal (noisy) or likely values.

In the following the invention will be explained for example with reference to the drawings. Fig. 1 shows schematically an arrangement according to the invention, Fig. 2 shows a

Arrangement according to FIG. 1 in detailed form.

According to FIG. 1, the overall system or the arrangement according to the invention comprises an automation system 2 comprising a multiplicity of units, which has a unit for ascertaining probability values which functions as determination unit 3 A and / or value ranges is bidirectionally connected. On the one hand, a user 1 can actuate units in the automation system 2; on the other hand, he also has the option of influencing the determination unit 3 and a filter system 4. The determination unit 3 is able to provide the user 1 with information about errors or deviations in the transmitter or sensor network of the automation system 2. This information can be output via any PC interface or directly via corresponding components in the automation system 2. Information from the automation system 2 can pass through the filter system 4 before being forwarded to the determination unit 3. Furthermore, a database 5 can provide the unfiltered values from the automation system 2 for later checks, for training or value modification in the determination unit 3 or for optimization activities and make them available to the filter system 4 or the determination unit 3.

Sensor or sensor networks or the automation system 2 comprise according to

Fig. 2 at least one or more sensor (s) 11 and actuator (s) 12, which sense the units not shown, and a communication unit 13 (cable, radio, etc.) between these units. In addition, control units 14 outputting output values can also be integrated into the sensor network 2 as additional sensors. The inventive arrangement for automatic detection and reporting of

Errors or deviations from normal values in a sensor network 2 furthermore comprise the determination unit 3, which recognizes relationships in the output values of the sensors by constructing a parameterized probability or learning model. This determination unit 3 can be subdivided into three modules: A learning management unit 15 gives the user the option of setting parameters of the

To influence learning algorithms or a change in the probability values or value ranges, which are used as the basis for the assessment of the initial values to achieve. For example, the user may decide which learning or modification model to use (if different learning methods are available), set parameters regarding the complexity of the model, determine maximum memory requirements, set the learning speed, etc. Further, the user has the option of Learning time to determine. Preferably, continuous learning is used because it allows the system to adapt automatically to changing environmental conditions. Nevertheless, the user has the opportunity to terminate the learning phase or temporarily suspend it. The learning management module 15 influences the learning algorithms 16 and the learning model 17. The term "learning" refers to the calculation of probability values and ₁ value ranges with the aid of distribution or probability laws or -modeilen (eg Gaussian models, histogram, Kalman filter, Markow chains or the. Like. And / or fuzzy logic method) understood.

The module learning model 17 contains the knowledge of the system in the form of a learned

5 model, ie a "summary" of all previous probabilities Based on this model, each new output value from the automation system is assigned a probability value and / or value range which indicates how normal or abnormal the new value is compared to the model ( "Überraschungsmaßzahl"). The more unusual the new starting value, the higher the value assigned to it. in ^{ι υ} If the learning model is given a lot of output values from the automation system, the model can not only make a statement about which output values are unusual, but it is also able to make a statement about which output values - in context with the other values

- would correspond to the normal case. As possible learning models come 15

Probability models (e.g., Gaussian models, histograms, Kalman filters,

Markow chains, etc.) are used.

When using probability models, the surprise measure of probability for a given model is:

20

where Sj is a sensor value, a _{j is} an actuator value, and M is the model. An actuator is understood to be a unit that can be influenced by control signals. Actuator values are, on the one hand, the control signals arriving at the actuator and, on the other hand, outgoing signals which describe the state of the actuator. In principle, actuator values and sensor values are to be regarded as identical output values.

In addition to probability models, other learning models such as fuzzy logic models can be used.

The surprising measure of the learning model 17 also has an influence on the filter system 4 having a learning filter 20, since this value decides whether a value is to be included or filtered out in the learning model. The learning system 20 receives new values either via the database 5 or from the automation system 2; From the learning model 17, the learning system 20 then receives an assessment (surprise measure) of these new values. If this surprise measure exceeds a certain threshold value, this value is not taken over into the learning model. Otherwise, the value is passed to the learning algorithms 16. 1 The results of the learning model 17 (information for the user) can be output either directly or by a controller unit 14 via corresponding output components 12 in the automation system 2. Furthermore, the results of the learning model can also be output via any PC interface 22.

5 The learning algorithm module 16 is responsible for taking values from the

Automation system 2, the filter system 4 or the database 5 receives to generate a learning model or adapt it. New values are not compared with a stored history of data, but directly with the data of the generated or existing learning model 17. The advantage of this approach is that a ^{ι υ} storage of collected data for the comparison is no longer necessary. This saves on the one hand the space requirement for the storage system, and on the other hand also improves the performance of the system, since the comparison is made only with the learning model, and not with a series of previous, stored data. Another

The advantage is that the privacy of the monitored users is maintained because 5 no data must be stored, but new data only with the

Learning model are compared.

The module learning algorithms 16 comprises the modules inference 18 and model adaptation 19. In the module inference 18, new values are compared with the learning model 17. The 0 result of this module is an assessment of whether or not a new baseline matches the learned model; in the case of non-compliance, a judgment is also made as to how the new baseline value is different from the normal case (for example, that it is 50% higher than the normal value). This assessment is done in collaboration with the learning model (17). 5 The following is an example of a Bayesian inference module. in the

Bayesian Inference Module, the system calculates this score as a probability of a value from a set of new values, for a given learning model and in context with the other values in that set:

P (Is ₁₁ S ₂ , ..., s _N , a ₁ , a ₂ , ..., a _R }; M) ⁰ (2)

sí a _s

where Si is a sensor value, a _j is an actuator value, M is the model, and 5

P (Js ₁₁ S ₂ , ..., s _N , ai, a ₂ , ..., a _R } | M) the input from the learning model 17, where a number N on

Sensors and a number R of actuators is used.

A second alternative is "Frequentist Inference." In this case, a confidence interval is calculated for an expectation value Reference to this confidence interval calculated. Again, other algorithms such as Fuzzy Logic are also possible here.

The model adaptation 19 is the part of the determination unit 3 or the module Lemalgorithmen 16, in which the model parameters are changed such that the normally observed values from the automation system 2 also represent the normal case in the learning model, ie in consequence in the module inference 18 a small surprise measure lead. The model adaptation module 19 changes the parameters of the learning model 17. In the case of probabilistic models, the model adaptation tries to influence or change the learning model so that the probability for new values from the automation system 2 is as high as possible. For this purpose, gradient-based optimization could be used: If a new value is transmitted to the determination unit, a parameter of the model can be changed as follows:

Mi «- Mi + ε 3 / 3Mi In (PKs ₁₁ S ₂ s _N , ai, a ₂ ΘR} | M)) (3)

where Mi is a parameter of the model M, and ε a parameter that influences the speed of learning (the part of the formula to the right of + is called the "Model Update term.") Further methods for this adaptation would be the "Newton's Method" or the " Expectation-Maximization Algorithm ".

The module learning algorithms 16 receives values either directly from the automation system 2, from a database 5 or from the filter system 4.

Another integral component is the filter system 4. It comprises the modules input filter 23 and learning system 20.

The input filter 23 is interposed between the data flow from the automation system 2 to the determination unit 3 and can be controlled by the user 1. From the automation system 2, the input filter 23 can receive data from the sensors 11, from the actuators 12 and also from the control units 14. The filtered values are then transferred to the module learning algorithms 16 of the determination unit 3. The purpose of this filter is to filter out nonstandard values known to the user. Examples of this would be if maintenance is performed on the heating, and the heating system would supply incorrect values during this time, or if the occupants are on holiday and the building or rooms are not used for some time, etc. The user has for the possibility to set different filter masks. These masks may be time-limited, they may include certain rooms, certain sensor areas (eg 1 temperature values), certain processes (eg holidays), or combinations of these masks. For example, a mask could include all temperature values in a particular room between 10:00 and 12:00. The crucial point in using these masks is that, while a certain part of the sensors

5 is filtered out, the rest is still doing its job. For example, if one selects the mask for vacation (with a certain time range), the system ignores the fact that the rooms and devices are not used for a long time - but at the same time it would detect intruders or other misconduct.

The learning system 20 is also the data flow from the automation system 2 to

'® determination unit 3 interposed, but works without interference by the user. Instead, it is controlled via the learning model 17. From the automation system 2, the learning system 20 may receive data from the sensors 11, from the actuators 12, from the database 5, and also from the control units 14. The

The task of the learning system 20 is to transfer values from the automation system only to the 5

Transfer determination unit 3, if the learning model should learn these values. When

The criterion for this information forwarding is the surprise measure used. The determination unit should only learn a new value if the value is not outside the norm. If the learning model 17 were also confronted with values Q which are far outside the norm, then the model parameters would have to be correspondingly strongly changed in order to integrate these values into the model. However, this would mean that the normal range of the learning model 17 would move away from the normal range in the normal world. The task of the learning system 20 is now to prevent this "negative influence" of the learning model 17. 5 As an alternative to blocking these initial values, the learning system can reduce the learning speed for unusual values, so that the number of times a value must occur until the learning model considers it to be normal, depending on the degree of surprise - the higher the surprise measure, the more often the value must occur.The surprise measure receives the learning system 20 from the learning model 17. If, for example, for the module learning algorithms 16 "gradient-based-method "and using the probability model from (Eq.1) for the learning model 17, the" model update term "in (Eq.3) would be very large if a very unlikely value is passed to the determination unit 3. In the sequel the parameters of the model are greatly changed In summary, it can be said that the learning system 20 5

Either filter out values with low probability or the parameter for the

Learning speed (Eq. 3) may decrease in relation to the probability. In addition to the filter system 4, a database 5 can be used. The

The database collects the unfiltered values from the automation system 2 for later checks, for the training of the determination unit 3, or for optimization activities, and can make them available to the filter system 4 or the determination unit 3. dates

5 from the automation system 2 can receive the database 5 from the sensors 11, from the actuators 12 and also from the control units 14. This stored information can be made available in the filter system 4 to both the input filter 23 and the learning system 20. Furthermore, the information can also be transferred to the learning model 17 of the determination unit 3. It is important in ^υ that this is unchanged values from the automation system. 2 The

Values are at least provided with a time stamp in this storage - the relationship between a particular unit of the automation system 2, the value from this unit, and the time of occurrence of this value must be maintained. The number of records stored depends on the desired use of this information. For this data storage, a database, but also any other form of storage system (e.g., lists, flat file, etc.) may be used.

The user 1 has various possibilities to interact with the arrangement,

20 e.g. via a unit 21. On the one hand, it can influence the automation system 2 by operating sensors 11 (for example, light switches, temperature controllers, etc.), or

Parameters of the sensors 11 or the controller units 14 (e.g., thresholds,

Time periods, etc.). This influence affects insofar as that this

Changes via the filter system 4 or passed directly to the determination unit 3 5. Furthermore, the user 1 has the opportunity to influence the learning management 15 and the input filter 23.

The invention further relates to a computer program product with corresponding software code sections or computer-readable program means with which the method according to the invention can be carried out on or with a computer.

According to the invention, probabilistic models or models equivalent thereto are used to calculate a probability or probability value indicating whether a sensor value or a combination of sensor values is common or unusual or surprising without previously being expected

Sensor value is calculated explicitly. The probability value is given with a 5

Threshold compared or it is checked whether the probability value in one

Probability range falls to unusual or exceptional sensor values or correlations or relationships between the measurements of individual Detect ₁ sensors. With this procedure, there is an improvement in the accuracy of judgment and prediction, in particular, of interrelations, a simplification of the installation, and an improvement in the operation and use of the monitored networks.

Thus, probabilistic models are used to assign a probability value to each sensor value or a number of sensor values without explicitly calculating a sensor value in value, and this probability value is used to detect network errors or deviations from normal operation and possible control and measurement Initiate rule actions.

As a result of the procedure according to the invention, the network or system is simplified or more easily controllable or controllable. Instead of having to define each threshold or range of values as a function of the type of sensor, because the predetermined value ranges or threshold values must be calculated based on actual sensor values or actual sensor value ranges, a calculation is carried out by means of probability values. For example, a pressure sensor has a completely different threshold compared to a temperature sensor, so that for each threshold individually a decision for the creation of the threshold must be made, by a specialist with accurate knowledge of the

2Q different measuring systems, sensors and determination of measured values. By contrast, by recourse to probability drawings and the resulting probability calculation, a comparison of actual or directly comparable values can be made. A probability value of 1: 10,000 means for different sensors, e.g. Pressure sensors and

25 temperature sensors always that this value can occur once under ten thousand values. Therefore, a single threshold or threshold range can be used for any sensor or any combination of sensors of the network, because each sensor provides a probability value. This results in considerable time and cost savings in the installation or construction of the network.

Furthermore, the system is simple to set up and operate. When the system displays an alarm, the reason for this alarm is quite clear, as the system provides a probability value; if it is an unjustified alarm, the threshold can easily be changed, e.g. such that the 35

The probability threshold for triggering this alarm is raised, because this alarm was judged to be inaccurate. A change of thresholds for actual or real measured values can be omitted.

Claims

P a t e. n t a n s p r e c h e:

5 1. Method for Monitoring or Determining the Deviation of Determined Values, or Output Values, ie, Measured Values and / or Status or State Values with Respect to Normal or Specified Values, the Values with Transmitters, in Particular State or Measured Value or status sensors, preferably sensors and / or data transmitters or actuators, in particular status or measured value or status transmitter networks, of ⁰ units detected or monitored therewith and determined in the form of output values, characterized in that

a probability value and / or value range for each encoder with the aid of distribution or probability laws or codes (eg Gaussian models, histograms, Kalman filters, Markov chains or the like and / or fuzzy logic methods) its next initial value, in particular for its course and / or the time of its occurrence and / or its magnitude and / or its change and / or its content, and / or

the output values of at least two different encoders are correlated with one another by using for these output values using distribution or

Probability laws or models, e.g. Gaussian models, histograms, Kalman filters, Markov chains or the like and / or fuzzy logic methods, a probability value and / or value range for their interdependence, in particular with regard to the time of their occurrence and / or their course 5 and / or its magnitude and / or its change and / or its content, and

it is ascertained for the output values of the encoders whether the probability value and / or value range of this output value lies outside the assigned probability range or deviates from the assigned probability value ⁰ .

2. The method according to claim 1, characterized in that upon detection of a deviation, in particular over a predetermined distance addition, a signaling is made or the delivery of a control signal is carried out with the particular 5 is attempted by the units a subsequent output value with a permissible

To receive value.

3. The method according to claim 1 or 2, characterized in that the individual donors associated or created probability value range or probability value is determined arbitrarily or determined based on previously obtained output values.

4. The method according to any one of claims 1 to 3, characterized in that the individual donors associated probability values or value ranges, in particular until the arrival of a new output value and recalculation of the probability values or value ranges, is stored.

5. The method according to any one of claims 1 to 4, characterized in that the individual donors associated probability value and / or the probability value range after the arrival of one, preferably each, new

Initial value, starting from or based on the previous for this

Transmitter determined probability value and / or

Probability value range are recalculated taking into account the new output value.

6. Method according to claim 5, characterized in that only newly arriving output values are used for the recalculation of the probability value and / or probability value range, which has a predetermined distance to the probability value range or probability value of the last output value with respect to course and / or arrival time and / or does not exceed magnitude and / or change and / or content and / or does not alter a present calculated probability value and / or probability value range beyond a predetermined amount and / or its present calculated probability values and / or probability values are within a predetermined range.

7. The method according to claim 1, wherein the permissible distance to the probability value range or probability value of the last output value and / or to a predefined calculated probability value and / or value range using probability models, eg gas models, histograms , Kaiman filters, Markow chains or the like. and / or fuzzy logic method is determined. ^

8. Arrangement for monitoring or determining the deviation of determined values or output values, i. of measured values and / or status or state values in relation to normal or setpoint values, the values being transmitted by encoders, in particular state or measured value or status transmitters, preferably sensors and / or data transmitters or

5 actuators, in particular measured value transducer networks (2), are detected by units detected or monitored thereby and delivered in the form of output values, in particular for carrying out the method according to one of claims 1 to 7, characterized

- that each of the encoders (2) is connected to a determination unit (3) for determining probability values and / or value ranges, in which

Using distribution or probability laws or models, e.g. Gaussian models, histograms, Kalman filters, Markov chains or the like and / or fuzzy logic methods, a probability value and / or value range for the course and / or the time of occurrence and / or the magnitude and / or change 5 and / or content of the next output value of the transmitter is determined, and / or

a determination unit (3) is provided for correlating the output values of at least two different encoders, in which, for these output values, 0 is used with the aid of distribution or probability laws or models, e.g. Gaussian models, histograms, Kalman filters, Markov chains or the like and / or fuzzy logic methods, a probability value and / or a

Probability range for their interdependence, in particular with regard to the date of their occurrence and / or their course and / or their magnitude and / or their change and / or their content are created, and

a discrimination or comparison unit or a learning system (20) is provided which checks a newly arriving initial value with regard to the course and / or time of its occurrence and / or its magnitude and / or its change and / or its content, this output value corresponds to the calculated probability value and / or falls within the calculated probability value range and, if appropriate, determines to what extent or which distance the probability value and / or value range of the received output value is remote from a predetermined or ascertained probability value and / or value range, and O

- That the detection unit (3) optionally a control unit for the units or a signaling unit is connected. ^

9. Arrangement according to claim 8, characterized

- That a filter system (4) is provided which sets a threshold, and

- That over the filter system (4) of the determination unit (3) for the determination of probability values and / or value ranges included learning model (17) only new

5 incoming initial values for recalculation of the probability values or value ranges are supplied, which fall below this threshold value, wherein the filter system (4), preferably the learning system (20) and an input filter (23).

10. Arrangement according to claim 8 or 9, characterized in that a ^{ι υ} Lernmanagement- or input unit (15) is provided with the

Probability values and / or probability value ranges can be changed and / or in particular for initialization, can be predetermined and / or given measured values can be added as additional output measured values of the determination unit (3) for the determination of

Probability values and / or value ranges can be fed. 5

11. Arrangement according to one of claims 8 to 10, characterized in that a learning model (17) and / or a memory unit (5) for a predetermined number of initial values obtained at certain times of the transmitter and / or 0 assigned to the individual transducers Probability values or value ranges and / or a module (16) are provided for learning algorithms, this module (16) having a module (18) for comparing new values with the learning model (17) and a module (18) for influencing the parameters of the Lem model (17).

12. Use of the method according to one of claims 1 to 8 and an arrangement according to one of claims 8 to 11, for determining the function and control of units of equipment and / or buildings and / or communication networks with respect to status or state values and / or measured values or parameters such as

Temperature, occupancy of rooms, stay of persons, electricity, water consumption, 0

Heating adjustment, blind position, lamp operations, or of units of

Industrial plants and / or vehicles with regard to measurement data, e.g. Tire pressure, braking behavior, engine speed, electronic parameters, personnel deployment, efficiency, etc., or units of distribution systems in terms of status or

State values and / or parameters such as e.g. Stock, consumption, sales, quantities, 5

Delivery quantity, etc.

Computer program product, in particular with program code means, for carrying out the method or all steps of any one of claims 1 to 7, if the program is running on a computer.

14. Computer program product with program code means which are stored on a computer-readable medium and cause the computer to perform the method according to any one of claims 1 to 7 when the program product is installed in a computer or an arrangement according to one of claims 8 to 14 or integrated.