DE102019208922A1 - Method and device for controlling a production process - Google Patents

Abstract

Method (10) for controlling a production process (20), characterized by the following features: - in the first process step (Mk) first measured values (xiMk) and in the second process step (Mj) second measured values (xiMj) are recorded (11), - with the First measured values (xiMk) and second measured values (xiMj), a probabilistic model is trained by unsupervised learning (12) and - on the basis of the model, quality defects occurring in the production process (20) are recognized on a case-by-case basis (13).

Description

The present invention relates to a method for controlling a production process. The present invention also relates to a corresponding device, a corresponding computer program and a corresponding storage medium.

State of the art

In the production industry, any standardized manufacturing method is referred to as a production process, with which, for example in an industrial company, a usable product is produced through mechanical or manual processing of raw materials or intermediate products. The end products, finished products, semi-finished products, intermediate products, materials or workpieces that are afflicted with quality defects in such a production process are referred to as defective production.

DE102009024101A1 discloses a method and control device for evaluating status data of a machine tool, the method comprising the following steps: the provision of the status data by an action recorder in the machine tool, the status data representing control commands or events provided in real time and the events being a reaction to the control commands represent in the machine tool, the transmission and storage of the status data from the action recorder in a logbook memory of a data processing unit independent of the action recorder and the retrieval and evaluation of the status data stored in the logbook memory by a monitoring unit that is located in a spatially remote position from the Data processing unit is arranged, the monitoring unit being designed to retrieve and output status data of different machine tools from a plurality of different logbook memories uvalue.

Disclosure of the invention

The invention provides a method for controlling a production process, a corresponding device, a corresponding computer program and a corresponding storage medium according to the independent claims.

The proposed method is based on the knowledge that a typical production chain is already equipped with different sensors or can be equipped with sensors of this type with reasonable effort. Their measured values can be used to train a predictive model that predicts product quality. The training of such a model is conventionally carried out by means of so-called supervised learning and therefore requires training data sets whose declared or target variables (labels) indicate the resulting quality of the product.

The approach described below takes into account the fact that there is sometimes no such database, especially since the latter is usually specific to one type of product. In addition, generic sensors are typically set up for regulation, but not for quality control. This means that even if labels are created, the measured values do not necessarily correlate with these labels. The existing sensors could rather be useless for quality control. For a data-based quality control according to the state of the art, additional sensors must therefore often be attached, e.g. B. cameras, acceleration or force sensors.

One embodiment of the invention is thus based on the insight that one hundred percent quality control using classical methods is generally not possible. In particular, checking the quality after each individual process step is particularly time-consuming and not economical. The individual machines are equipped with sensors, but these are designed for regulation and not for quality control. The measured values do not necessarily allow a statement about the product quality. In addition, it is often not economical to provide the data with quality labels, as this can be time-consuming, requires expert knowledge and sometimes even incorrect productions.

Against this background, the proposed solution is based on a manufacturing process in which several process steps are connected in series or in parallel. The machines or the process steps use different sensors for the control. The following explanations describe how to use the existing sensor data to identify possible deviations from the normal production behavior of the process steps. The method can determine which process step deviates from the intended normal behavior and thus endangers the overall quality. In addition, a probability is calculated for each individual product, which indicates individual products that may be defective and will be rejected.

One advantage of this solution is that existing sensors can be used for quality control to control individual process steps or production machines. The method used for this does not require any of Experts classified training data sets. They can therefore also be used in production, where, according to the state of the art, only the quality of a very small amount of the end products can be checked. A quality control of individual process steps is typically not provided here, but is also made possible by the approach presented, which in turn allows an indication of problematic individual process steps.

While known methods require a direct correlation between measured values and quality and do not take into account the dependencies between processes, the approach presented below also works if the individual sensor data do not correlate with the quality. Additional sensors, as they are sometimes required in the context of other approaches to data-based quality control, can be dispensed with. Even for new products, a new model does not necessarily have to be trained as long as they go through the same process chain.

The measures listed in the dependent claims enable advantageous developments and improvements of the basic idea specified in the independent claim.

Figure list

• 1 das Flussdiagramm eines Verfahrens gemäß einer ersten Ausführungsform.
• 2 das Diagramm einer Produktionsprozesskette mit verschiedenen Prozessschritten.
• 3 schematisch ein Steuergerät gemäß einer zweiten Ausführungsform.

Exemplary embodiments of the invention are shown in the drawings and explained in more detail in the description below. It shows:

• 1 the flow chart of a method according to a first embodiment.
• 2 the diagram of a production process chain with various process steps.
• 3 schematically a control device according to a second embodiment.

Embodiments of the invention

A basic idea of the proposed method ( 10 ) consists in using historical sensor data to use machine learning (ML) to train a model to evaluate the dependencies between the process steps. Using such models, the measured values delivered by the sensors in individual process steps are to be predicted.

The deviations between the prediction and the actual measured value are then used to sort out individual defective products or to initiate an emergency shutdown (hereinafter referred to as “emergency shutdown”).

The starting point of the following considerations is a process chain consisting of several production steps, of which at least two process steps are monitored with sensors, for example for the purpose of control. A layout is also known that describes how the process steps depend on one another. A further prerequisite is that measurement data recorded in the individual process steps are available and communication is possible at least between adjacent process steps so that sensor data and predictions can be passed on.

1 Based on these assumptions, illustrates the sequence of the proposed procedure ( 10 ) for quality control: After collecting historical data (process 11 ) a probabilistic model is trained with the measured values through unsupervised learning (process 12th ), on the basis of which quality defects that occur in the production process can be identified on a case-by-case basis (process 13 ).

In 2 is an example of a diagram of such a production process ( 20th ) with different process steps (M ₁ , M ₂ , M ₃ ) that are connected, for example, in series or in parallel, without excluding other layouts.

bezeichnet.For the purpose of regulation, the process steps (M ₁ , M ₂ , M ₃ ) are monitored by sensors. The measured values for the product i in the process step M _j , which statistically form a multidimensional time series of variable length, are included below $x_i^{{\mathrm{M.}}_j}$

designated.

der im Prozessschritt M_j erfassten Messwerte $x_i^{M_j}$

wird als mit θ parametrisierte Funktion f_θ der im Prozessschritt M_k erfassten Messwerte $x_i^{M_k}$

modelliert. Anhand von historischen Daten wird sodann der Parameter θ optimiert, was später im Einzelnen beschrieben sei.For the in 2 In pairs, connected by an arrow and in this sense “neighboring” pairs (M ₁ , M ₂ ) and (M ₂ , M ₃ ) of process steps, a probabilistic model is trained to determine how the measured values recorded in process step M _j differ from the depend on measured values recorded in process step M _k . The conditional probability $$\text{p}\left({\text{x}}_{\text{i}}^{{\text{M.}}_{\text{j}}}|{\text{x}}_{\text{i}}^{{\text{M.}}_{\text{k}}}\right)={\text{f}}_{\mathrm{\theta }}\left({\text{x}}_{\text{i}}^{{\text{M.}}_{\text{k}}}\right)$$

of the measured values recorded in process step M _j $x_i^{{\mathrm{M.}}_j}$

is as f θ with parameterized function of _{θ k} in process step M acquired measurements $x_i^{{\mathrm{M.}}_k}$

modeled. The parameter θ is then optimized on the basis of historical data, which will be described in detail later.

entsprechen. Mit hoher Wahrscheinlichkeit ist dieser Normalfall gegeben. Bei zu geringer Wahrscheinlichkeit wird das gerade gefertigte Produkt auf geeignete Weise markiert und später zum Zwecke einer Nachkontrolle aussortiert. Wenn im Prozessschritt M_j mehrere Mängel auftreten, kann - wie eingangs beschrieben - ein Not-Aus bewirkt oder eine anderweitige Maßnahme zur Qualitätssicherung getroffen werden.The model that has been trained once is used during ongoing production operations as follows: The Control of M _j can execute the learned model. In real time, it receives the measured values recorded in process step M _k and compares whether the new measured values correspond to the predicted distribution $p\left(x_i^{{\mathrm{M.}}_j}|x_i^{{\mathrm{M.}}_k}\right)=f_{\mathrm{\theta }}\left(x_i^{{\mathrm{M.}}_k}\right)$

correspond. This normal case is very likely. If the probability is too low, the product that has just been manufactured is appropriately marked and later sorted out for the purpose of a follow-up check. If several defects occur in process step M _j , an emergency stop can - as described above - be effected or another quality assurance measure can be taken.

von unterschiedlicher Länge sein können. $f_{\mathrm{\theta }}\left(x_i^{M_k}\right)$

kann hierbei durch ein sogenanntes Sequenz-zu-Sequenz-Modell (sequence-to-sequence model) angenähert werden, deren Eingangssequenz die Zeitreihe $x_i^{M_k}$

darstellt. Ein vorzugsweise bidirektionales gestapeltes (stacked) rückgekoppeltes oder rekurrentes neuronales Netz (RNN), vorzugsweise mit getasteten rekurrenten Einheiten (gated recurrent units, GRUs) und sogenanntem langen Kurzzeitgedächtnis (long short-term memory, LSTM), fasst hierbei als Leser (encoder) die in Zeitreihe $x_i^{M_k}$

erfassten Messwerte zusammen.In principle, a single model is sufficient for this. Nevertheless, the effectiveness of the procedure ( 10 ) can be improved by modeling the dependencies between several neighboring process steps individually. For each pair of process steps (M _k , M _j ) for which a model is to be trained according to the above scheme, historical measurement data are required, with the time series $x_i^{{\mathrm{M.}}_j}\text{and}{x}_i^{{\mathrm{M.}}_k}$

can be of different lengths. $f_{\mathrm{\theta }}\left(x_i^{{\mathrm{M.}}_k}\right)$

can be approximated by a so-called sequence-to-sequence model, the input sequence of which is the time series $x_i^{{\mathrm{M.}}_k}$

represents. A preferably bidirectional stacked feedback or recurrent neural network (RNN), preferably with keyed recurrent units (gated recurrent units, GRUs) and so-called long short-term memory (LSTM), includes the reader (encoder) in time series $x_i^{{\mathrm{M.}}_k}$

recorded measured values together.

auch lediglich das auf entsprechende Weise kodierte Signal an die Steuerung von M_k weitergegeben werden.In production, instead of raw data $x_i^{{\mathrm{M.}}_k}$

also only the signal encoded in a corresponding manner can be passed on to the control of M _k .

The signal encoded in this way is then decoded again with a writer (decoder) which also has a recurrent architecture. A mean value and a variance are predicted for each point in time; the model is therefore probabilistic. Its output is a probability distribution. The model is trained on this by optimizing a suitable scoring rule (scoring function) or plausibility function (likelihood function).

This model architecture works equally when x is one-dimensional and when there are multiple types of sensors. If N _j indicates the dimensionality of the measured values in process step M _j , then for a given pair of process steps either a single model corresponding to the described architecture with multiple inputs and multiple outputs (multiple input multiple output), or N _k · N _j one-dimensional models are trained. In addition, models for other combinations (M _k , M _j ) of process steps can be trained.

User-adjustable parameters include the probability thresholds used to decide which product is rejected and the way in which the various models and their predictions are combined to initiate an emergency stop. Exemplary user settings could provide, for example, that the relevant component or the end product is sorted out if the measurement data of the production process ( 20th ) deviate from the forecast by more than 20%. It could also be provided that if more than 10% of the processed components are sorted out in a process step, an emergency stop is initiated and the user is warned that the process step is causing quality problems.

This method ( 10 ) can be in software or hardware, for example, or in a mixed form of software and hardware, for example in a control unit ( 30th ) be implemented as shown in the schematic representation of the 3 clarified.

QUOTES INCLUDED IN THE DESCRIPTION

This list of the documents listed by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Patent literature cited

• DE 102009024101 A1 [0003]

Claims (10)

Method (10) for controlling a production process (20) with a first process step (M _k ) and a second process step (M _j ) dependent on the first process step (M _k ), characterized by the following features: in the first process step (M _k ) first readings $\left(x_i^{{\mathrm{M.}}_k}\right)$

and second measured values in the second process step (M _j ) $\left(x_i^{{\mathrm{M.}}_j}\right)$

recorded (11), - with the first measured values $\left(x_i^{{\mathrm{M.}}_k}\right)$

and second measured values $\left(x_i^{{\mathrm{M.}}_j}\right)$

a probabilistic model is trained through unsupervised learning (12) and - on the basis of the model, quality defects occurring in the production process (20) are recognized on a case-by-case basis in an ongoing production operation (13). Method (10) according to Claim 1 , characterized by the following features: - the model is trained on it (12), a conditional probability distribution of the second measured values $\left(x_i^{{\mathrm{M.}}_j}\right),$

given the first readings $\left(x_i^{{\mathrm{M.}}_k}\right),$

to calculate and - the recognition (13) of the quality defects takes place, if the second measured values $\left(x_i^{{\mathrm{M.}}_j}\right)$

fall below a predetermined minimum probability according to the probability distribution. Method (10) according to Claim 1 or 2 , characterized by the following feature: - as soon as a specified production error rate is exceeded due to the quality defects, an emergency shutdown of the production plant is initiated. Method (10) according to one of the Claims 1 to 3 , characterized in that the first measured values $\left(x_i^{{\mathrm{M.}}_k}\right)$

or the second measured values $\left(x_i^{{\mathrm{M.}}_j}\right)$

relate to at least one of the following measured variables: an electrical current, an electrical energy, a temperature, an axis position or an operating state of a machine used in the production process (20). Method (10) according to one of the Claims 1 to 4th , characterized by the following feature: the model is a preferably stacked or bidirectional recurrent neural network, in particular with a long short-term memory or keyed recurrent units. Production process (20), characterized by the following feature: - the production process (20) comprises a method (10) according to one of the Claims 1 to 6th . Production process (20) Claim 6 , characterized by at least one of the following features: - the first process step (M _k ) is based on the first measured values $\left(x_i^{{\mathrm{M.}}_k}\right)$

regulated or - the second process step (M _j ) is based on the second measured values $\left(x_i^{{\mathrm{M.}}_j}\right)$

regulated. Computer program which is set up, the method (10) according to one of the Claims 1 to 7th execute. Machine-readable storage medium on which the computer program is based Claim 8 is stored. Device (30) which is set up, the method (10) according to one of the Claims 1 to 7th execute.

Images