DE29700028U1 - Self-diagnosing device with fuzzy neuro logic diagnostic device - Google Patents

Description

GR 96 Q 5163

Description

Self-diagnostic device with fuzzy neuro-logic diagnostic facility
5

The invention relates to a device (in particular a device with rotating parts that are driven by a shaft) that has its own fault diagnosis device. The fault diagnosis is carried out by evaluating operating variables using fuzzy logic.

Vibration problems and mechanical damage to equipment with rotating parts, such as large pumps, make efficient monitoring systems for the early detection of mechanical changes appear sensible. To determine the measurement data required for diagnostic purposes, the following are preferably suitable:

- the speed of the shaft driving the rotating parts 20

- the characteristic values derived from the oscillation path of the wave, measured in two directions offset by 90 ^a from each other, such as the magnitude and phase of the first to the fifth harmonic oscillation, the magnitude and frequency of the whirl, as well as the spectral residual energy

- the magnitude and phase of the first to fifth harmonic oscillations, as well as the spectral residual energy, which can be derived from the vibration velocities on the housing, measured in two directions offset by 90 ^{° to} each other.

To diagnose malfunctions in a rotating device, it is necessary to compare the measured data with a parameter. The GDR industrial patent DD 221 550 Al specifies a method for determining the diagnostic parameter. The parameter is defined as a

GR 96 G 5163

The mean value of all signal components in the broadband frequency range is calculated using a specific formula.

In recent times, it has been preferred to evaluate machine parameters using fuzzy rules. It is known that a fuzzy system is transferred to a neural network and then transferred back after a learning process. This learning process adapts the parameters of the fuzzy system to the machine parameters. The disadvantage of this approach is that the fuzzy must be readjusted each time the device's parameters change according to the operating conditions. It also seems impossible to transfer the diagnostic device's fuzzy system to other devices of the same type.

The object of the invention is to provide a diagnostic and condition monitoring device based on fuzzy logic for such a device. This object is achieved by the features of claim 1. Preferred developments are specified in the subclaims.

In particular, the device according to the invention has measuring devices, which are attached to the shaft, for example, and record its speed and its vibration path. The measurement data obtained by the measuring device are transferred to an online diagnostic and monitoring system for machine diagnostic and monitoring purposes. The diagnostic and monitoring system has an evaluation and standardization system, based on fuzzy neurologic. The normal values of the device on which the fuzzy system is based are created by transferring the machine measurement values and the fuzzy system to a neural network. There, normal values are created for the machine parameters and the fuzzy system is subjected to a learning process. Following the learning process, the fuzzy system is transferred back with the normal values. By comparing the current measurement values and the normal values, the diagnostic system receives information.

GR 96 G 5163

indicates deviations of the device from the normal state. If the normal state of the device is changed, for example by modifying the device, a learning process is carried out by a neural network, through which the fuzzy system is adapted to the new normal state values. This adaptation is achieved by a second evaluation system, based on the original normal values, examining whether and in what combination these deviate from the current normal state.

The fuzzy control system can be optimized by a neural network optimizing the fuzzy rules of the second evaluation system by selecting from given rules those that were applied during device operation.

Even greater control and monitoring reliability can be achieved by determining the vibration speed of the housing using measuring sensors in addition to controlling the shaft. This measurement data can then be processed in the diagnostic device using fuzzy rules.

A device with rotating parts that is equipped with such a fault diagnosis system has in particular the following advantages:

A simple and direct implementation of expert knowledge in fuzzy rules at an abstract level into the diagnostic system is possible.
30

The expert knowledge implemented in the system adapts itself to the specific device through independent learning.

The knowledge required for diagnosis is provided in a rule base using fuzzy logic. The rules

GR 96 G 5163

written in as general a form as possible so that the knowledge can be transferred to the same or similar devices.

In order to adapt the abstract rule base to the specific device, a self-learning system based on a neural network is used. For this purpose, the fuzzy logic rules are integrated into a suitable topography of a neural
Network and adapted to the device in question using training patterns. After the learning process in the neural network, a reverse transformation into a fuzzy logic rule base is carried out.

The neural network serves exclusively to adapt the
general fuzzy logic rule bases to the specific device and is not a direct part of the fuzzy logic system.

Fault diagnosis using fuzzy logic therefore involves two sub-steps:

- In the first step, the standard behavior of the device component monitored by measurement technology is determined. This standard behavior
is highly dependent on various operating parameters. A separate rule base represents the standard behavior.

In the second step, the relationships between the properties
of the characteristics and possible errors. Because the
Dependencies of operating parameters in the normal state

0 status rule base, the rules can be set up in a generalized form. These rule bases
conclude from the deviations of the parameters from their
Normal behavior for possible errors.

For this purpose, it is necessary that the diagnostic system

machine condition is modelled. To describe the machine condition, the properties size, variance and trend of the

GR 96 G 5163

considered characteristic values are used. The properties trend and variance of the characteristic values are not treated as dependent on operating parameters and are therefore only available once for each characteristic value as an input variable. They are used directly in the rule base for error diagnosis.

The trend parameter is described as a linguistic variable using several membership functions. It is characterized by a regression line, the slope of which indicates to the evaluation system that there are impending changes in the device status. This can result in an error message that can be output as a linguistic variable and described by further membership functions. These membership functions can transform the fuzzy value into a possible measure of the occurrence of an error.

Since even identical machines exhibit strong differences in their vibration behavior under the same operating conditions, it is necessary that the general rule bases adapt to the specific devices. To do this, all possible rules are set up between the linguistic variables and assigned a weight. This formal fuzzy logic rule base is transformed into a network topology.

The rule conclusions and rule weights are adjusted through a multi-stage learning process with corresponding learning data. If the neural network has a method for eliminating unused weights between the neurons, the unused rules can be removed. This creates a fuzzy logic control system after the back transformation to describe the characteristic value depending on the operating parameter.

A fuzzy neuro transformation can also be used to adapt the general diagnostic rule base to the 5 specific conditions of a machine. In principle, if secure data material is available for individual

GR 96 G 5163

Error conditions make an automated adjustment of the rule base possible.

The following is an example of a large pump equipped with a fault diagnosis system whose function is based on fuzzy neurologic.

Show

FIG 1 shows the schematic structure of the diagnosis using a fuzzy neuro system;

FIG 2 the membership functions of the linguistic variable "speed";
15

FIG 3 the membership functions of the linguistic variable "size of a characteristic value" at the output;

FIG 4 the membership functions for the linguistic variable "trend of a characteristic value";

FIG 5 the membership functions for the linguistic variable "variance of a characteristic value";

FIG 6 the membership functions of the linguistic variable "state of the machine";

FIG 7 the membership functions of the linguistic variable "Error N";
30

FIG 8 Example of fault diagnosis using fuzzy rules;

FIG 9 Test data for learning process in the neural network;

FIG 10 Fuzzy logic rule base modified with neural network;

GR 96 G 5163

FIG 11 shows modelling of the parameter dependence according to Figure 9, implemented using an adapted fuzzy system.

In the example, the only parameter to be considered is the working speed of the machine. In machines with variable operating speed, this usually has a dominant influence on the vibration parameters. The dependence of the other parameters on operating parameters is probable, but will be omitted for the sake of a simpler presentation.

The shaft of the large pump is equipped with a speed sensor, which transmits its data to the diagnostic system. The structure of the diagnostic system is shown schematically in Figure 1.

One possibility for modeling the operating parameter dependency, here the speed, is shown below:

The range of the working speed of the machine - e.g. 400 to

2 000 revolutions per minute - is divided into a linguistic variable (LV) with 12 membership functions (ZF) (Figure 2). For this purpose, the speed range is divided into ten speed groups. These are:
25

"too small" - lower limit

"1" to "10" - e.g. ten speed groups "too large" - upper limit.

The input range of each linguistic variable "characteristic value of a characteristic value" at the input is divided into individual and overlapping membership functions analogous to the speed (Figure 2). The membership functions cover the entire value range evenly. The speed groups therefore receive a characteristic value of a certain size:

GR 96 G 5163

"too small" - measured value below the sensor resolution "B 01" to "B 10" - e.g. ten characteristic value groups "too large" - measured value above the end of the measuring range or outside of physical sense.
5

The property "size of each characteristic value" at the output is formally described with a linguistic variable with five membership functions (Figure 3). Membership functions are assigned to the characteristic values, which evaluate them qualitatively. These membership functions are denoted by

"too small" - smaller than the lower limit "small" - measured value smaller than the normal range - "normal" - measured value in the normal range "large" - measured value larger than the normal range "too large" - larger than the upper limit.

With these prerequisites, the following rules can be formally established to describe the dependence of the characteristic values on an operating parameter, where each characteristic value represents a speed range:

Rulel

5 OPTIONS

WEIGHT=IOO
IF (KennGrlE IS too_small) AND {Speed IS one) THEN

KennGrl = too_small
END
30

Rule2
OPTIONS

WEIGHT=IOO
IF {KennGrlE IS B_01) AND (speed IS one) THEN 5 KennGrl = small

END

GR 96 G 5163

Rule3 OPTIONS

WEIGHT=50

IF (KennGrlE IS B_01) AND (Speed IS one) THEN KennGrl = normal END

RuI e4 OPTIONS WEIGHT=IOO IF (KennGrlE IS B_02) AND (Drehzahl IS ein) THEN

KennGrl = normal END

Rule5 OPTIONS

WEIGHT=50 IF (KennGrlE IS B_03) AND (Speed IS one) THEN KennGrl = normal END

Rule6 OPTIONS WEIGHT=2 IF (KennGrlE IS B_04) AND (Drehzahl IS ein) THEN

KennGrl = capital END

Rule7 0 OPTIONS

WEIGHT=IO IF (KennGrlE IS B_04) AND (Speed IS one) THEN

KennGrl = large END 35

GR 96 G 5163

R\ile8 OPTIONS

WEIGHT=20

IF (KennGrlE IS B_05) AND {speed IS one) THEN KennGrl = large END

Rule 9 OPTIONS WEIGHT=30 IF (KennGrlE IS B_06) AND (Drehzahl IS eins) THEN

KennGrl = large END

RuIel0 OPTIONS

WEIGHT=40 IF (KennGrlE IS B_07) AND (Speed IS one) THEN KennGrl = large 0 END

Rulell OPTIONS WEIGHT=50 5 IF (KennGrlE IS B_08) AND (Drehzahl IS ein) THEN

KennGrl = large END

Rule2 OPTIONS

WEIGHT=GO IF {KennGrlE IS B_09) AND (RPM IS one) THEN

KennGrl = large END 35

GR 96 G 516:

• · * &igr;

Rulel3
OPTIONS

WEIGHT=7 0

IF (KennGrlE IS B_10) AND (Speed IS one) THEN KennGrl = large
END

Rulel4
OPTIONS
WEIGHT=IOO

IF (KennGrlE IS too_large) AND (RPM IS one) THEN

KennGrl = too_large
END

These rules are used to assign a specific input value of the characteristic value to a linguistic variable with an abstract value designation, e.g. "normal" (see Rule4). The rule weight is used to smoothly transition from several input ranges to one output range (see Rule7 to Rule13). Transitions between different input ranges to different output ranges can also be carried out using the rule weights (see Rule2 to Rule6).

5 The dependencies of operating parameters are thus modeled in the rules with their weights. For each membership function of each linguistic variable "size of a characteristic value" at the input, the number of rules to be set up is at least equal to the membership function of the linguistic variable "speed". To describe the machine state, the properties size, variance and trend of the characteristic values considered are used.

The properties trend and variance of the characteristic values are not treated as dependent on operating parameters and are therefore only available once as an input variable for each characteristic value.

GR 96 G 5163

They are therefore used directly in the rule base for error diagnosis as shown in Figure 1.

The trend parameter is described as a linguistic variable using five membership functions (Figure 4). Increases in the regression line by ± 0 are considered a "constant" membership function and are therefore considered unremarkable. Increases up to ± 1 are evaluated using the "increasing" or "decreasing" membership functions and can be considered as an indication of impending changes. Increases over ± 1 are to be interpreted as strong changes in the parameter and are described using the "strongly increasing" or "strongly falling" membership functions.

The variance of a parameter (Figure 5) can be considered as a standardized value between zero and one, where zero means no variance and one means large variance.

The state of the device is divided into five membership functions (Figure 6) and denoted by:

"no statement" - default range if no rule applies "normal" - normal range in speed group N "suspicious" - one or more parameters have left their normal range and are in the "small" or "large" range

"out of range" - one or more parameters have left their normal range and are in the range "too small ¹ ' or "too large"

- "invalid above" - area is currently not used.

The description of the linguistic variable "error N" for the M errors considered (Figure 7) is described in a simple form by two membership functions "unlikely" and "probable". The defuzzified value can thus be used as a possibilistic measure for the occurrence of the error.

GR 96 G 5163

13

It is therefore possible to establish formal rules to describe the machine condition and the individual errors considered in the following way:

Rulel

IF {KennGrl IS large) AND (KennGrl_Trend IS rising) THEN

Error = probably END
10

Rule2

IF (KennGr2 IS small) AND (KennGr2_Trend IS constant) AND (KennGr2_Variance IS constant) THEN

Error = unlikely END

Rule3

IF (KennGrl IS normal) AND {(KennGrl_Trend IS strongly_increasing) OR {KennGrl_Trend IS strongly_decreasing)) THEN

Errorl = probably Error3 = probably END
25

The concrete formulation of the rule base takes place in the development project of the system (e.g. Figure 8). During the learning and testing phase in a field test, the rule base is modified with regard to the parameters used and the rules established.

The general rule bases adapt to the specific device. The rule base for modeling the operating parameter dependency on the device is modified using a neural network. For this purpose, all possible rules between the linguistic variables "size of a characteristic value" are formally defined at the input of the linguistic variable

5R 96 G 5163

"Speed" and the linguistic variable "Size of a characteristic value" at the output and assigned the weight 100 (see the following RuIeXX).

RuIeXX
OPTIONS

WEIGHT=IOO
IF (KennGRlE IS "ZFE") AND {Drehzahl IS "ZFD") THEN KennGrl = normal.
END

with

„ZFE" = „too_small", „B_01",.., „B_10", „too_big" „ZFD" = „too_small", „one",.,, „ten", „too_big"

This formal fuzzy logic rule base is transformed into a network topography. Through a multi-stage learning process with corresponding learning data (e.g. Figure 9), the rule conclusions and the rule weights are adapted {Figure 10).

Since the neural network has a procedure for eliminating unused weights between neurons, the unused rules (weight = 0) can be removed.

This creates, after the reverse transformation, a fuzzy logic control system to describe the characteristic value depending on the operating parameter (Figure 11).

A fuzzy neuro transformation is also used to adapt the general diagnostic rule base to the specific conditions of a device. In principle, if secure data material is available for individual error states, e.g. from simulation calculations or tests, an automated adjustment of the rule base is conceivable (Figure 8).

Claims (6)

GR 96 G 5163 15 Protection claims 1. Device with an electronic evaluation device in which a fuzzy logic for evaluating the operating variables of the device is implanted, characterized by measuring devices arranged on the device for generating measured values, an online connection of the measuring devices to the electronic evaluation device, a first system implanted in the electronic evaluation device for standardizing the measured values, a second system implanted in the electronic evaluation device which evaluates the standardized measured values as operating variables of the device according to a system of fuzzy rules simulating the device, and an output unit fed by the second implanted system for outputting diagnostic and error detection signals. 2. Device with evaluation device according to claim 1, characterized in that the first 0 implanted system contains fuzzy rules which model the normal behavior of the device. 3. Device with evaluation device according to claim 1 or 2, characterized in that the fuzzy rules of the second system evaluate the deviation of the current measured values from the normal values. 4. Device with evaluation device according to claim 2 or 3, characterized by a neural network implanted in the evaluation device, the topology of which corresponds to the fuzzy rules and subjects the first and/or second system to a learning process. GR 9ζ G 5163 5. Device with evaluation device according to one of claims 1 to A, characterized in that the device contains rotating parts and the measuring devices record at least characteristic data from the vibration spectrum of the rotating parts. 6. Device with evaluation device according to claim 1, characterized in that the device is a large-scale device.