DE102019216999A1 - Device and method for monitoring a machine - Google Patents

Abstract

Device (100) and method for monitoring a machine (114), wherein in a work sequence of the machine (114) for a product type or a class for a component measurement data (112) for a parameter of the machine (114) or for a parameter of one of its Components are recorded, wherein in a first phase a first model (106) for a first distribution of the measurement data (112) is determined depending on a plurality of such work processes, wherein in a second phase a second model (108) is dependent on a plurality of such work processes. for a second distribution of the measurement data (112), a result of the monitoring being determined as a function of a comparison of the first model (106) with the second model (108) as a function of a measure for the difference between the models or a change in the measure .

Description

State of the art

The highest possible machine availability is aimed for in order to achieve productivity goals. In order to achieve this goal, unplanned machine downtimes should be avoided as far as possible.

Disclosure of the invention

This is made possible by the subject matter of the independent claims.

A method for monitoring a machine provides that measured data for a parameter of the machine or for a parameter of one of its components are recorded in a work sequence of the machine for a component, wherein in a first phase a first model for a The first distribution of the measurement data is determined, with a second model for a second distribution of the measurement data being determined in a second phase depending on a plurality of such work processes, with a result of the monitoring depending on a comparison of the first model with the second model depending on a measure for the diversity of the models or a change in dimension is determined. The distributions each represent a probability measure for values of the parameter in a work sequence of the machine for the component to which the recorded measurement data are assigned. This approach is used instead of operating hours or time-dependent planned maintenance intervals, instead of maintenance in the event of specific suspicions that can be identified by recognizable wear or a loss of quality, and instead of condition-based maintenance based on machine data or suitable additional sensors. In the first phase, the first model is created as a good model from the measurement data of the parameter of the machine or one of its components. In the second phase, the first model is created as an actual model from the measurement data of the parameter of the machine or one of its components. A deviation of the actual model from the good model makes it possible to reliably identify an irregularity. In this way, in particular, different loads are taken into account that occur due to different processing programs, for the production of variants or other products, or due to changed machine settings, without additional adaptation effort of the evaluation logic.

Measurement data are preferably determined for various components, the first distribution and / or the second distribution being determined as a function of the measurement data for the various components. Different components require different workflows. This is taken into account by using the measurement data from the various components.

The measure for the difference is preferably determined as a function of a Kullback-Leibler divergence or a Jensen-Shannon divergence.

In one aspect, the measurement data are defined by an energy consumption of the machine per component for the workflow, and / or the measurement data are defined by an energy consumption of a component of the machine per component for the workflow. The energy consumption is a particularly well-suited parameter that reflects wear on the machine or individual components of the machine particularly well for monitoring purposes.

In one aspect, standardized measurement data are determined for the first distribution and / or the second distribution as a function of the measurement data and a deterministic model for the workflow or as a function of a probabilistic graphical model for the workflow. This normalization removes influences that arise, for example, from override settings. This increases the robustness of the monitoring.

In this aspect, at least one machine parameter is preferably recorded, the deterministic model or the probabilistic graphical model being determined as a function of the at least one machine parameter. The consideration of machine parameters relevant to energy consumption, such as an override setting of a machine tool, also improves the model, since parameters set by a machine operator that have an impact on the comparison result are taken into account.

The deterministic model or the probabilistic graphical model is preferably determined depending on the measurement data of the component from the same workflow. The influences can thus be compensated particularly well.

The first phase is preferably carried out with a new machine or when the machine is started up, the second phase being carried out in operation of the machine following the first phase. The good model is created with a machine that is as good as new. In this way, an undesired deviation can be identified particularly well during later operation.

Preferably, depending on the measure for the difference, a The probability of failure of the machine is determined and / or a signal is output as a function thereof. This enables feedback to the machine operator or a machine control system. This allows immediate consequences to be taken.

The first distribution and / or the second distribution can be determined for work processes with different machine settings. Models with these distributions offer a determination of a standardized wear parameter across different manufactured product types and specific machine settings without having to carry out measurements outside of productive operation.

The first model and the second model are preferably a parametric model, a kernel density estimate model or a flow-based generative model. These models are particularly suitable because they provide a likelihood that can be evaluated in comparison.

A device for monitoring a machine comprises at least one processor, a memory for instructions that can be executed by the at least one processor, a first model, a second model, an input for measurement data from a machine and an output for a signal, the device being designed that To execute the method when the at least one processor executes the instructions.

• 1 eine Vorrichtung zur Überwachung einer Maschine,
• 2 ein Verfahren zur Überwachung der Maschine,
• 3 Aspekte eines Modells zur Überwachung der Maschine.

Further advantageous embodiments emerge from the following description and the drawing. In the drawing shows:

• 1 a device for monitoring a machine,
• 2 a method for monitoring the machine,
• 3 Aspects of a model for monitoring the machine.

In 1 is a device 100 shown.

The device 100 comprises at least one processor 102 . The device 100 includes a memory 104 for at least one processor 102 executable instructions. The processor 102 and the memory 104 can be part of a machine control system.

The device 100 includes a first model 106 and a second model 108 . In the example, the models are in memory 104 saved.

The device 100 includes an entrance 110 for measurement data 112 from a machine 114 and an exit 116 for a signal 118 . The entrance 110 is trained in a work flow of the machine 114 for a component, measurement data for a parameter of the machine 114 or to record one of its components for a parameter. The measurement data 112 can include an identification of the component that is machined or manufactured in the machine's workflow. The measurement data 112 may include machine settings that will operate the machine while the measurement data is being acquired. The measurement data 112 can include at least one machine parameter that is relevant for the machine's work flow 114 is set when the component is manufactured or machined.

The device 100 can be connected to the machine control system of the machine 114 for recording machine parameters, type identification and energy consumption, and possibly additional sensors for recording energy consumption.

The device 100 is designed to carry out the method described below when the at least one processor 102 executes the instructions.

The procedure for monitoring the machine 114 is based on the 2 described.

In one step 202 are in a work flow of the machine 114 For a product type or a class for a component, measurement data is recorded for a parameter of the machine or for a parameter of one of its components. A class for a component comprises a product type for a component and variants thereof. A variant of a component is similar to the component with regard to the geometric type or the nature of the material, so that, for example, essentially the same parameter can be expected. The measurement data can be due to an energy consumption of the machine 114 be defined for each component for the workflow. The measurement data can be due to the energy consumption of a component of the machine 114 be defined for each component for the workflow. It can be provided that at least one machine parameter is recorded.

In one step 204 the measurement data are assigned to an identification of the product type or the class for the component. The identification makes it possible to evaluate the measurement data for each component in the following steps.

These steps 202 and 204 can be performed repeatedly. This will be in the steps 202 Measurement data determined for different components and in steps 204 their respective identification assigned.

In a first phase 206 becomes the first model depending on a plurality of such work processes 106 intended for a first distribution of the measurement data.

The first phase 206 is in the example with a new machine 114 or when the machine is commissioned 114 carried out.

The first distribution is determined, for example, as a function of the measurement data of the various components. This is described in more detail below.

The first distribution can be determined based on identical workflows or for workflows with different machine settings. The workflows differ due to the different machine settings.

For the first distribution, normalized measurement data can be determined depending on the measurement data and a deterministic model for the workflow or depending on a probabilistic graphical model for the workflow.

If it is provided that at least one machine parameter is recorded, the deterministic model or the probabilistic graphical model is determined in one aspect as a function of the at least one machine parameter.

The deterministic model or the probabilistic graphical model can be determined from the same workflow depending on the measurement data of the component. In this example, a work process is a specific temporal sequence of spatial movement of the machine 114 , the part or the components of the machine 114 defined, which is specified by the machine control for manufacturing or processing the component.

The first model 106 can be a parametric model or a kernel density estimate model or a flow-based generative model. The first model 106 represents a good model in the example. An example for creating the model 106 is described in more detail below.

In a second phase 208 the second model becomes dependent on a plurality of such work processes 108 intended for a second distribution of the measurement data.

The second phase 208 is in the example on the first phase 206 subsequent operation of the machine 114 carried out.

The second distribution is determined, for example, as a function of the measurement data from various components.

The second distribution can be determined for identical work processes or for work processes with different machine settings.

For the second distribution, normalized measurement data can be determined depending on the measurement data and a deterministic model for the workflow or depending on a probabilistic graphical model for the workflow.

If it is provided that at least one machine parameter is recorded, the deterministic model or the probabilistic graphical model is determined in one aspect as a function of the at least one machine parameter.

The deterministic model or the probabilistic graphical model can be determined from the same workflow depending on the measurement data of the component.

The second model 108 can be a parametric model, or a kernel density estimate model, or a flow-based generative model. The second model 108 represents an actual model in the example. Details are described below.

In one step 210 becomes a result of the monitoring dependent on a comparison of the first model 106 with the second model 108 determined depending on a measure for the diversity of the models or a change in the measure.

The measure of the difference is determined, for example, as a function of a Kullback-Leibler divergence or a Jensen-Shannon divergence based on the two distributions. In the example, the good model is compared with the actual model learned in this way. Details are described below.

In one step 212 a failure probability of the machine becomes dependent on the measure for the difference 114 certainly. The measure for the difference characterizes a characteristic value, for example wear of the machine 114 or one of its components. A failure probability can also be determined, with which preventive maintenance work can be planned.

In one step 214 a signal can be output depending on the result of the comparison. In the example, the result of the comparison characterizes wear on the machine 114 . With known wear limits, for example Visualize a current machine status for a machine operator or a service team.

For example, a traffic light representation of the current machine status is output on a display.

This simple approach to visualizing the machine state can be used to establish a normal state of the machine 114 green, a deteriorating condition of the machine 144 yellow and a critical condition of the machine 114 highlighted in red. The classification can be based on the relative change in the result of the comparison without having to know absolute limit values.

For example, a percentage representation of the wear can be used as a wear indicator for the machine 114 respectively. With this form of representation, information on the probability of failure can also be conveyed, so that maintenance work can be planned. For this purpose, absolute limit values are used for a characteristic value that is determined depending on the result of the comparison. In the example, limit values are defined both for good status and for an actual failure.

In the following, a solution approach is described that is based on the assumption that there is a measurable variable that causes wear on the machine 114 or individual components of this machine 114 can capture. The machine is in the concrete implementation 114 for example a machine tool. The measurement data characterize the electrical energy consumed by the machine 114 or a number N of individual components of this machine 114 such as drive motors.

3 represents aspects of a model 300 that can be used for the approach. In the example are the input variables of the model 300 a total energy consumption of the machine 114 per component 302 , an energy consumption of each of the considered components 1 to N of the machine 114 per component 304 , at least one machine parameter 306 and a type part number 308 , as identification of the component. The model 300 provides a distribution in this example 310 on the electrical energy consumed when manufacturing a large number of components. The model 300 can be the first model 106 , ie in the example as a good model, or as a second model 108 , ie as the actual model in the example.

When creating the good model used in the example, it is assumed that the machine 114 is in good condition and it makes sense to compare the actual model with the good model from this condition. For example, under operational conditions, several components are manufactured without an override setting, and the measurement data indicating the electrical energy consumed by the machine 114 for each of the components, stored together with the identification of the components.

These measurement data are used to calculate the good model via typical distributions p (e | t, p, v ₀ ) of absorbed electrical energy e ein
Component t with parameters p at normalized speed v _{0 of} a movement of the machine 114 or to learn one of its components.

Depending on the data availability and the complexity of the distribution, one of the models described is used as a good model. With these models, a likelihood can be determined that can be evaluated.

• - parametrisches Modell p_θ(e|t,p,v₀) bestimmt, sofern für die Verteilung ein bekanntes parametrisches Modell hierfür vorhanden ist,
• - Kernel-Density Estimate p_KDE(e|t,p,v₀) bestimmt, sofern kein parametrisches Modell bekannt ist, aber eine Annahmen, z.B. smoothness, gängiger Kernel hierbei passen, oder
• - Flow-based generative Model p_Flow(e|t,p,v₀) bestimmt, sofern kein parametrisches Modell bekannt ist, aber angenommen werden kann, dass das erste Modell 106 sehr komplex ist.

For example, the Gut model is called

• - parametric model p _θ (e | t, p, v ₀ ) is determined if a known parametric model is available for the distribution,
• - Determines the kernel density estimate p _KDE (e | t, p, v ₀ ) if no parametric model is known, but assumptions, e.g. smoothness, common kernels, or
• - Flow-based generative model p _Flow (e | t, p, v ₀ ) determined if no parametric model is known, but it can be assumed that the first model 106 is very complex.

In the example, the measurement data required for this are recorded and evaluated during the manufacture of the respective component t. In this case, t represents the identification of the component.

After the good model has been created, the same measurement data is recorded to create the actual model during the production of further components in the example. For a component t, the measurement data together with the identification of this
Component t recorded and evaluated. The machine's override setting is optionally available as a machine parameter 114 recorded and evaluated.

During operation, the example continues to measure the measurement data that indicate the electrical energy consumed by the machine 114 for each of the components included. In the example, the measurement data are recorded and stored for each component with the associated identification of the respective component.

Für einen Vergleich von Gut-Modell und Ist-Modell ist es im Beispiel sinnvoll, die aufgenommene elektrische Energie zu normieren. Dadurch können Einflüsse, die sich aus unterschiedlichen Arbeitsabläufen oder Maschinenparametern, wie Override-Parametern, die die Override-Einstellung definieren, ergeben, eliminiert werden. Für die Normierung kann ein Energie-Modell verwendet werden. Override-Parameter könne Einfluss auf die aufgenommene Energie haben, z.B. da in einem Ausbohren eine durch den Override-Parameter veränderte Geschwindigkeit eine andere Energie erfordert, als dies bei der Erstellung des Gut-Modells ohne Override-Parameter der Fall war. Das Energie-Modell ist beispielsweise eine Abbildung, die bei einem Aufbohren eine Schnittleistung proportional zu einer Drehzahl der Maschine 114 abbildet. Gegebenenfalls ist die Abbildung mit einem Stumpfungsfaktor skaliert.In the example, the last N measurements are saved for the last N components. In the example, the number N depends on the needs of the model to be learned. The following applies to the number N in the example $$N_{\theta }\le N_{K\mathrm{D.}\mathrm{E.}}\le N_{\mathrm{F.}lOw}$$

For a comparison of the good model and the actual model, it makes sense in the example to normalize the electrical energy consumed. In this way, influences that result from different work processes or machine parameters, such as override parameters that define the override setting, can be eliminated. An energy model can be used for normalization. Override parameters can have an influence on the energy consumed, for example, because a speed changed by the override parameter in a boring operation requires a different energy than was the case when the good model was created without override parameters. The energy model is, for example, an illustration that, when boring, shows a cutting power proportional to the speed of the machine 114 maps. If necessary, the illustration is scaled with a blunting factor.

Alternatively, a Probabilistic Graphical Model, PGM, can be trained, which is used for normalization. In the example, the input variables of the PGM are the total energy consumption of the machine 114 per component 302 , the energy consumption of each of the considered components 1 to N of the machine 114 per component 304 , the machine parameter 306 and a type part number 308 , as identification of the component. The PGM 300 in this example supplies the normalized speed v ₀ . Thus, the good model or the actual model becomes an energy consumption standardized with regard to the speed 310 determined per component.

Since the energy consumption for the production of a component can also fluctuate, a distribution via energy consumption can also be used as a characteristic of the state of the machine 114 be used. There can also be a change in this distribution as an indicator of a change in the state of the machine 114 be used.

Bei einer Messung der beschriebenen Messdaten sind im Beispiel die Werte von T, P, V und E vorhanden. Dadurch lässt sich der additive Einfluss des Rauschens ε identifizieren, eine andere angenommene Geschwindigkeit v vorgeben und das Rauschen ε wieder addieren.For example, a functional relationship between the speed v changed by the override parameter and the energy e can be reduced to a linear combination of speed-dependent and component-dependent basic functions and additive noise ε. The noise ε is assumed to be normally distributed, for example. This allows the PGM to be modeled using the following equation: $$e={\theta }^T\left(v,t,p\right)f_{\varphi }\left(t,p\right)+\epsilon mit\epsilon \sim N\left(\mathrm{0,}\sigma \right)$$

When measuring the measurement data described, the values of T, P, V and E are available in the example. This allows the additive influence of the noise ε to be identified, a different assumed speed v to be specified and the noise ε to be added again.

The actual model is determined analogously to the good model from the last N standardized measurements as a model q (e | t, p, v ₀ ).

In the example, the models are trained to map the input variables to output variables that correspond to the mapping available in the measurement data. For this training, a parameter optimization with a gradient descent method is carried out for parametric models, preferably as a function of an error measure that is defined by a deviation of the result of the mapping by the model from the mapping specified from the measurement data. In the example, the model is trained when the machine is new 114 , since a good case can be assumed here. If this is not possible, for example due to a later installation, the degree of change in the characteristic value supplied by the actual model is already meaningful.

Diese sind nicht symmetrisch und haben unterschiedliche Eigenschaften. Die Verteilungen sind identisch, wenn gilt $$D_{KL}\left(p\left|\right|q\right)=D_{KL}\left(p\left|\right|q\right)=0$$

Für die Kullback-Leibler-Divergenz $$D_{KL}\left(p\left|\right|q\right)={\displaystyle {\int }_{e}p\left(e\right)log\frac{p\left(e\right)}{q\left(e\right)}}de$$

treten an Stellen große Werte auf, an denen die Verteilung q sehr geringe Wahrscheinlichkeitswerte und die Verteilung p demgegenüber höhere Wahrscheinlichkeiten aufweist.The comparison of the actual model with the good model takes place in the example by determining the respective Kullback-Leibler divergences $$\begin{array}{l}{\mathrm{D.}}_{K\mathrm{L.}}\left(p\left|\right|q\right)\\ {\mathrm{D.}}_{K\mathrm{L.}}\left(q\left|\right|p\right)\end{array}$$

These are not symmetrical and have different properties. The distributions are identical if applies $${\mathrm{D.}}_{K\mathrm{L.}}\left(p\left|\right|q\right)={\mathrm{D.}}_{K\mathrm{L.}}\left(p\left|\right|q\right)=0$$

For the Kullback-Leibler divergence $${\mathrm{D.}}_{K\mathrm{L.}}\left(p\left|\right|q\right)={\displaystyle {\int }_{e}p\left(e\right)lOG\frac{p\left(e\right)}{q\left(e\right)}}de$$

large values occur at points where the distribution q has very low probability values and the distribution p, on the other hand, has higher probabilities.

treten an Stellen große Werte auf, wenn neue Moden in der Verteilung q auftreten, die in der Verteilung p noch nicht vorhanden waren. Beispielsweise führt eine Drift in der Verteilung zu höheren Energien.For the Kullback-Leibler divergence $${\mathrm{D.}}_{K\mathrm{L.}}\left(q\left|\right|p\right)={\displaystyle {\int }_{}q\left(e\right)lOG}\frac{q\left(e\right)}{p\left(e\right)}de$$

large values occur at points when new modes appear in the distribution q that were not yet present in the distribution p. For example, a drift in the distribution leads to higher energies.

mit $$m=\frac{1}{2}\left(p+q\right)$$

Der Vergleich von Gut-Modell und Ist-Modell liefert in diesem Beispiel einen Kennwert für den Verschleißzustand der Maschine 114 oder deren Verschleißrelevanter Komponenten.A combined solution using Jensen-Shannon divergence would also be possible: $${\mathrm{D.}}_{K\mathrm{L.}}\left(p\left|\right|m\right)+{\mathrm{D.}}_{K\mathrm{L.}}\left(q\left|\right|m\right)$$

With $$m=\frac{1}{2}\left(p+q\right)$$

In this example, the comparison of the good model and the actual model provides a characteristic value for the state of wear of the machine 114 or their wear-relevant components.

Claims (14)

Method for monitoring a machine (114), characterized in that in a work sequence of the machine (114) for a product type or a class for a component, measurement data (112) for a parameter of the machine (114) or for a parameter of one of its components is recorded (202), wherein in a first phase (206) depending on a plurality of such work processes, a first model (106) for a first distribution of the measurement data (112) is determined, in a second phase (208) depending on a plurality of such work processes Workflows a second model (108) is determined for a second distribution of the measurement data (112), a result of the monitoring depending on a comparison of the first model (106) with the second model (108) depending on a measure of the difference between the models or a change in the measure is determined (210). Procedure according to Claim 1 , Characterized in characterized in that measurement data for various components is determined (202), said first distribution and / or the second distribution is determined depending on the measurement data of the various components (206, 208). Method according to one of the preceding claims, characterized in that the measure for the difference is determined as a function of a Kullback-Leibler divergence or a Jensen-Shannon divergence (210). Method according to one of the preceding claims, characterized in that the measurement data are defined by an energy consumption of the machine per component for the workflow, and / or that the measurement data are defined by an energy consumption of a component of the machine per component for the workflow. Method according to one of the preceding claims, characterized in that standardized measurement data are determined for the first distribution and / or the second distribution depending on the measurement data and a deterministic model for the workflow or depending on a probabilistic graphical model for the workflow (206, 208 ). Procedure according to Claim 5 , characterized in that at least one machine parameter is recorded (202), the deterministic model or the probabilistic graphical model being determined (206, 208) as a function of the at least one machine parameter. Procedure according to Claim 5 or 6th , characterized in that the deterministic model or the probabilistic graphical model is determined as a function of measurement data of the component from the same workflow (206, 208). Method according to one of the preceding claims, characterized in that the first phase (206) is carried out with a new machine or when the machine is started up, the second phase (208) in an operation of the machine following the first phase (206) is carried out. Method according to one of the preceding claims, characterized in that depending on the measure for the difference, a failure probability of the machine is determined (212) and / or a signal is output as a function thereof (214). Method according to one of the preceding claims, characterized in that the first distribution and / or the second distribution are determined for work processes with different machine settings (206, 208). Method according to one of the preceding claims, characterized in that the first model and the second model are a parametric model, or are a kernel density estimate model, or are a flow-based generative model. Device (100) for monitoring a machine (114), characterized in that the device (100) has at least one processor (102), a memory (104) for instructions that can be executed by the at least one processor (102), a first model (106) , a second model (108), an input (110) for measurement data (112) from a machine (114) and an output (116) for a signal (118), the device (100) being designed according to the method one of the Claims 1 to 11 execute when the at least one processor (102) executes the instructions. Computer program, characterized in that the computer program is computer-readable Includes instructions, when they are executed by a computer, the method according to one of the Claims 1 to 11 expires. Computer program product, characterized in that the computer program product comprises a computer-readable medium on which the computer program after Claim 13 is stored.

Images