DE102023114025A1 - Method for detecting a clogged differential pressure line of a pressure measuring arrangement - Google Patents

Abstract

A computer-implemented method for detecting a blocked differential pressure line of a pressure measuring arrangement, comprising the following steps:- detecting a pressure measurement signal by means of the pressure measuring arrangement (1), preferably for a defined period of time;- combining, preferably successive measured values of the pressure measurement signal into, preferably overlapping blocks (Bt-1, Bt), which preferably each have the same length;- generating, in particular, acoustic frequency spectra for each block, wherein each frequency spectrum has a plurality of frequency components (c1,t-1; cn,t);- entering the, in particular acoustic frequency spectra of the blocks into a neural network which is configured to detect a blocked differential pressure line based on the frequency spectra of the blocks;- generating an output which indicates whether or not a blocked differential pressure line is present.

Description

The invention relates to a computer-implemented method for detecting a clogged differential pressure line of a pressure measuring arrangement, a data processing device with means for carrying out the method, a computer program, a computer-readable medium and a pressure measuring arrangement.

The European patent EP 1 840 549 B1 protects a device for detecting a blockage of differential pressure lines, the device comprising two pressure sensors for detecting time series of the static pressure in one differential pressure line each and a differential pressure sensor for detecting time series of the difference between the pressures in the two differential pressure lines and three calculation units for calculating the fluctuations of the two static pressures and the differential pressure based on the time series and the square sums of the fluctuations; correlation coefficient calculation units for determining the correlation coefficients between the time series of a static pressure and the differential pressure and an evaluation unit which detects blockages in the differential pressure lines based on the correlations and identifies which of the differential pressure lines is or are blocked.

The German patent application DE 10 2013 110 059 A1 describes a differential pressure measuring arrangement which detects clogged differential pressure lines based on a correlation of a differential pressure signal, which represents a pressure difference between a first medium pressure and a second medium pressure, and a temperature signal, which correlates with a temperature of a differential pressure line.

Furthermore, there are other approaches to detecting a clogged differential pressure line. In particular, there are approaches in which the method is based on changing the frequency of the pressure signal fluctuations.

All approaches are useful in themselves, but each has specific weaknesses. Some methods do not work at all or only work to a limited extent if the underlying process is too dynamic and subject to large fluctuations. Other methods only work in a certain pressure signal range that is regularly left by dynamic processes. Still other methods have weaknesses if resonances arise due to long differential pressure lines, which affect the signal power and can sometimes artificially increase the signal power.

The invention is based on the object of remedying this situation.

The object is achieved according to the invention by the method for detecting a blocked differential pressure line according to patent claim 1, the method for training a neural network according to patent claim 2, the data processing device according to patent claim 14, the computer program according to patent claim 15, the computer-readable medium according to patent claim 16 and the pressure measuring arrangement according to patent claim 17.

• - Erfassen eines Druckmesssignals mittels der Druckmessanordnung, vorzugsweise für einen definierten Zeitraum;
• - Zusammenfassen von, vorzugsweise aufeinanderfolgenden Messwerten des Druckmesssignals zu, vorzugsweise überlappenden Blöcken, die vorzugsweise jeweils die gleiche Länge aufweisen;
• - Erzeugen von, insb. akustischen Frequenzspektren für jeden Block, wobei jedes Frequenzspektrum mehrere Frequenzkomponenten aufweist;
• - Eingabe der, insb. akustischen Frequenzspektren der Blöcke in ein neuronales Netzwerk, welches dazu konfiguriert, insb. trainiert ist, eine verstopfte Wirkdruckleitung anhand der Frequenzspektren der Blöcke zu erkennen;
• - Erzeugen einer Ausgabe, die angibt, ob eine verstopfte Wirkdruckleitung vorliegt oder nicht.

The computer-implemented method according to the invention for detecting a clogged differential pressure line of a pressure measuring arrangement comprises the following steps:

• - detecting a pressure measurement signal by means of the pressure measurement arrangement, preferably for a defined period of time;
• - combining, preferably consecutive, measured values of the pressure measurement signal into, preferably overlapping blocks, which preferably each have the same length;
• - generating, in particular, acoustic frequency spectra for each block, wherein each frequency spectrum has several frequency components;
• - Input of the, in particular acoustic frequency spectra of the blocks into a neural network which is configured, in particular trained, to detect a blocked differential pressure line based on the frequency spectra of the blocks;
• - Generate an output indicating whether or not a clogged differential pressure line is present.

• - Erfassen eines Druckmesssignals mittels der Druckmessanordnung, vorzugsweise für einen definierten Zeitraum;
• - Zusammenfassen von, vorzugsweise aufeinanderfolgenden Messwerten des Druckmesssignals zu, vorzugsweise überlappenden Blöcken, die vorzugsweise jeweils die gleiche Länge aufweisen;
• - Erzeugen von, insb. akustischen Frequenzspektren für jeden Block, wobei jedes Frequenzspektrum mehrere Frequenzkomponenten aufweist;
• - Eingabe der, insb. akustischen Frequenzspektren der Blöcke in ein neuronales Netzwerk zum Trainieren des neuronalen Netzwerkes anhand der erzeugten Frequenzspektren der Blöcke.

The invention further relates to a computer-implemented method for training a neural network for detecting a clogged differential pressure line of a pressure measuring arrangement, comprising the following steps:

• - detecting a pressure measurement signal by means of the pressure measurement arrangement, preferably for a defined period of time;
• - combining, preferably consecutive, measured values of the pressure measurement signal into, preferably overlapping blocks, which preferably each have the same length;
• - generating, in particular, acoustic frequency spectra for each block, wherein each frequency spectrum has several frequency components;
• - Input of the, especially acoustic frequency spectra of the blocks into a neural network to train the neural network using the generated frequency spectra of the blocks.

According to the invention, the recording and evaluation of a frequency spectrum of a pressure measurement signal from a pressure measuring arrangement is proposed. The frequency spectrum can preferably be an acoustic frequency spectrum which is in a range of less than 20 kHz, in particular less than 2.5 kHz, in particular in the range of 0-2 kHz, and very particularly in the range of 5 Hz - 200 Hz. The (acoustic) frequency spectrum characterizes the pressure measuring arrangement and is essentially a fingerprint of the pressure sensor. A change in the pressure measuring arrangement affects the frequency spectrum, in particular the acoustic frequency spectrum. This is used according to the invention to train a neural network to recognize clogged differential pressure lines.

An advantageous embodiment of the method according to the invention provides that, before the frequency spectra of the blocks are input into a neural network, the frequency spectra of several, preferably consecutive blocks are first aggregated to form a two-dimensional spectrogram, which has a time on the abscissa and an amplitude of the Fourier coefficients on the ordinate, and the spectrograms are used as input for the neural network. In particular, the embodiment can provide that a power is calculated for each frequency component for the spectrograms.

A further advantageous embodiment of the method according to the invention provides that for the training of the neural network, a binary label is assigned to each spectrogram, which indicates whether a clogged differential pressure line is present or not.

A further advantageous embodiment of the method according to the invention provides that each spectrogram is processed by the neural network through several, preferably two convolution layers and/or is reduced in dimension by a max-pooling layer.

A further advantageous embodiment of the method according to the invention provides that the neural network is a convolutional network which is set up to use convolutional layers, in particular depth-separable convolutions. In particular, the embodiment can provide that flattening is carried out in the convolutional network, in which the multidimensional output data structure of a respective convolutional layer of the convolutional network is converted into a one-dimensional output data structure.

A further advantageous embodiment of the method according to the invention provides that a multi-stage, preferably a three-stage classifier made up of fully connected layers of decreasing size, generates a scalar output, in particular from the output data structure, which is mapped into a value range [0; 1], preferably by a sigmoid activation function, in order to generate the output that indicates whether a clogged differential pressure line is present or not.

A further advantageous embodiment of the method according to the invention provides that the frequency spectra are generated by means of a discrete Fourier transformation, in particular a fast Fourier transformation.

A further advantageous embodiment of the method according to the invention provides that the frequency spectra of the respective blocks are high-pass filtered so that frequency components that are below a predeterminable cutoff frequency are essentially filtered out or attenuated. In particular, the embodiment can provide that at least one variance is calculated for each block to filter out the idle phases of the process, and the variance is used to decide whether it is a idle phase of the process. In particular, the embodiment can provide that the variance is compared with a configurable or adjustable threshold value in order to decide whether it is a idle phase of the process, wherein the threshold value is preferably in a range ≤ -70 dB, particularly preferably ≤ -80 dB, very particularly preferably in the range from -80 to -90 dB.

The invention further relates to a data processing device with means for carrying out the method for detecting a blocked differential pressure line and/or for carrying out the method for training a neural network according to one of the previously described embodiments.

The invention further relates to a computer program with instructions which, when the program is executed by a computer, cause the computer to carry out the method for detecting a clogged differential pressure line and/or the method for carrying out the method for training a neural network according to one of the previously described embodiments.

The invention further relates to a computer-readable medium containing instructions which, when executed by a computer, cause the computer to carry out the method for detecting a clogged differential pressure line and/or the method for carrying out the method for training a neural network according to one of the previously described embodiments.

• - einen Druckmessumformer zum Erfassen zumindest eines Mediendrucks eines Mediums;
• - eine Betriebs- und/oder Auswerteschaltung zum Bereitstellen eines Druckmesssignals, welches von dem zumindest einen Mediendruck abhängt;
• - zumindest eine Wirkdruckleitung, welche an einen Druckeingang des Druckmessumformers angeschlossen ist, um dem Druckmessumformer mit dem zumindest einen Mediendruck zu beaufschlagen;
• - eine Detektionseinheit, welche vorzugsweise als Teil der Betriebs- und/oder Auswerteschaltung ausgebildet und dazu eingerichtet ist, das Verfahren zur Erkennung einer verstopften Wirkdruckleitung und/oder das Verfahren zur Durchführung des Verfahrens zum Trainieren eines neuronalen Netzwerkes nach einer der zuvor beschriebenen Ausführungsformen auszuführen.

The invention further relates to a pressure measuring arrangement, at least comprising:

• - a pressure transducer for detecting at least one media pressure of a medium;
• - an operating and/or evaluation circuit for providing a pressure measurement signal which depends on the at least one media pressure;
• - at least one differential pressure line which is connected to a pressure inlet of the pressure transducer in order to apply the at least one media pressure to the pressure transducer;
• - a detection unit, which is preferably designed as part of the operating and/or evaluation circuit and is set up to carry out the method for detecting a blocked differential pressure line and/or the method for carrying out the method for training a neural network according to one of the previously described embodiments.

• 1: eine Gesamtdarstellung eines Ausführungsbeispiels einer erfindungsgemäßen Druckmessanordnung,
• 2: ein Sensormodul und ein elektrisch mit dem Sensormodul verbundenes Elektronikmodul der Druckmessanordnung im Detail,
• 3: ein allgemeines Blockdiagramm des computer-implementierten Verfahrens zum Erkennen von verstopften Wirkdruckleitungen einer Druckmessanordnung,
• 4: ein weiteres Blockdiagramm des Verfahrens zur Erkennung einer verstopften Wirkdruckleitung einer Druckmessanordnung, in dem sowohl die Lern- als auch die Inferenzphase dargestellt ist,
• 5: ein weiteres Blockdiagramm des Verfahrens zur Erkennung einer verstopften Wirkdruckleitung einer Druckmessanordnung mittels eines neuronalen Netzwerkes, und
• 6: ein Spektrogramm für mittelwertbereinigte und mittels einer diskreten Fouriertransformation in den Frequenzraum transformierter Chunks.

The invention is explained in more detail with reference to the following drawings. It shows:

• 1 : an overall view of an embodiment of a pressure measuring arrangement according to the invention,
• 2 : a sensor module and an electronic module of the pressure measuring arrangement electrically connected to the sensor module in detail,
• 3 : a general block diagram of the computer-implemented method for detecting clogged differential pressure lines of a pressure measuring arrangement,
• 4 : another block diagram of the method for detecting a clogged differential pressure line of a pressure measuring arrangement, showing both the learning and the inference phase,
• 5 : another block diagram of the method for detecting a clogged differential pressure line of a pressure measuring arrangement by means of a neural network, and
• 6 : a spectrogram for chunks adjusted for mean values and transformed into frequency space using a discrete Fourier transform.

1 shows a simplified and schematic overall view of a pressure measuring arrangement. In 1 a differential pressure measuring arrangement is shown. Although described below using the differential pressure measuring arrangement, the invention is not limited to this, but can also be transferred to any other pressure measuring arrangement, in particular an absolute and/or relative pressure measuring arrangement. The 1 The differential pressure measuring arrangement shown comprises a differential pressure transmitter 10, which has a sensor module 11, which is arranged between a first, high-pressure side process connection flange 12 and a second, low-pressure side process connection flange 13, and an electronic module 14. Both modules are arranged in a housing of the differential pressure transmitter. The electronic module can be connected to a process control system 18 via a two-wire line 16, wherein the electronic module 14 can communicate and be supplied with energy via the two-wire line. The two-wire line can be operated in particular as a field bus according to the Profibus or Foundation Fieldus standard or according to the HART standard. Such differential pressure transmitters are known per se and are manufactured and marketed by the applicant, for example, under the Deltabar brand.

The differential pressure measuring arrangement can further comprise a differential pressure sensor 20 for installation in a pipeline 21. The differential pressure sensor comprises an orifice 22, a first pressure tap channel 23 on a high-pressure side of the orifice 22 and a second pressure tap channel 24 on a low-pressure side of the orifice 22. The high-pressure side process connection flange 12 is connected to the high-pressure side pressure tap channel 23 via a high-pressure side differential pressure line 25, and the low-pressure side process connection flange is connected to the low-pressure side pressure tap channel 24 via a low-pressure side differential pressure line 26. The terms “high-pressure side” and “low-pressure side” refer to a pressure difference generated by a flow (in the drawing from left to right), which is proportional to the square of the flow rate and is, for example, in the order of magnitude of a few 10 to 100 mbar. The static pressure on which this flow-dependent pressure difference is superimposed can, for example, range from 1 bar to several 100 bar.

2 shows a sensor module 11 and a Electronic module 14 in detail. The arrangement of the two 2 For reasons of simplicity, the modules shown are arranged side by side and in this respect different from 1 The sensor module in this case is a piezoresistive pressure sensor element 110 with a measuring membrane 112, wherein the pressure sensor element 110 has resistance elements in a bridge circuit in order to convert a pressure-dependent deformation or deflection of the measuring membrane into an electrical sensor module signal. Instead of the piezoresistive pressure sensor, a capacitive pressure sensor can also be provided, in which case the measuring membrane has an electrode whose capacitance compared to an electrode on a rigid counter body is a measure of the pressure-dependent deformation or deflection of the measuring membrane. Details of electrical converters are familiar to a person skilled in the field of pressure measurement technology and do not need to be explained in more detail here. The sensor module outputs the sensor module signal, which depends on the detected pressure difference, to the electronic module 14, wherein an operating and/or evaluation circuit 160 of the electronic module 14 generates a differential pressure measurement signal representing the pressure difference based on the sensor module signal and can output it, for example, via the two-wire line 16, in particular to the process control system 18. For this purpose, the operating and/or evaluation circuit can have a computing unit, for example in the form of a microprocessor 162 for processing sensor module signals of the pressure sensor 110 digitized by means of an ADC 164.

The pressure measuring arrangement further comprises a detection unit 150 which is designed to carry out the computer-implemented method described below. The detection unit 150 can, as in 2 shown, be part of the operating and/or evaluation circuit 160 or alternatively be implemented separately as a separate unit, e.g. in a cloud.

The computer-implemented method is divided into several method steps, a preprocessor step executed in a preprocessor 152 (cf. 3 ), in which the pressure measurement signal or differential pressure measurement signal is transformed into a frequency spectrum, a feature extractor 154 (cf. 4 ) in which features are derived from the frequency spectrum using a neural network. Furthermore, an optional classifier step can be carried out in a classifier 156 in which a decision is made on the basis of the previously derived features as to whether the differential pressure line 25, 26 is blocked or not. The method steps are carried out one after the other. 3 shows a general block diagram of the computer-implemented method for detecting clogged differential pressure lines 25, 26 of a pressure measuring arrangement 1.

In the first step, the pressure measurement signal is first sampled at the highest possible sampling rate in order to obtain successive measured values, preferably digitized measured values, particularly preferably measured values digitized by the ADC 164. The sampling rate is preferably selected such that it is at least a few 10 Hz, preferably at least 50 Hz, particularly preferably at least 100 Hz.

In order to prevent idle phases belonging to the process, the pressure values of which are determined using the pressure measuring arrangement 1, from making it difficult to detect a blocked effective pressure line 25, 26, it is necessary to detect and filter out these idle phases as quickly as possible. This can basically be done in different ways. A simple subsequent method can provide for the preferably consecutive measured values to be combined in the preprocessor step into preferably overlapping blocks B _t-1 , B _t (hereinafter also referred to as chunks), preferably of the same length. Chunks B _t-1 , B _t each form inseparable units and all measured values of a chunk B _t-1 , B _t are each processed together. Then, in the preprocessor step, a variance and optionally a mean value are calculated for each chunk B _t-1 , B _t . The variance is then compared with a configurable or adjustable threshold value. If the variance of the chunk B _t-1 or B _t is below the threshold value, the chunk B _t-1 or B _t is ignored because it is highly likely that it is a rest phase in the process. The threshold value depends on the process and can vary from process to process. For example, the threshold value for a level process in which a level is determined using a pressure measuring arrangement can be different than for a flow process in which a flow is determined using a differential pressure measuring arrangement. However, the threshold value is usually in a range ≤ -70 dB, preferably ≤ -80 dB, particularly preferably in the range of - 80 to -90 dB.

The remaining chunks B _t-1 , B _t are then mean-corrected in the preprocessor step and transformed into the frequency domain using a discrete Fourier transformation, e.g. a Fast Fourier Transformation (FFT), in order to obtain the Fourier coefficients for each chunk B _t-1 , B _t required for the subsequent feature extractor step. 6 shows an example of such a spectrogram. After the Fourier transformation, optionally in the preprocessor step before It can be seen that a high-pass filter 158 is used, which operates at a predeterminable cutoff frequency in order to filter out the actual pressure value. The predeterminable cutoff frequency can, for example, be selected so that it is in the range of a few Hz, preferably less than 5 Hz, particularly preferably in the range of 1-5 Hz. The output of the preprocessor after processing consists of the (possibly high-pass filtered) Fourier coefficients of a chunk.

In the feature extractor step following the preprocessor step, the features required to detect a clogged differential pressure line 25, 26 are determined. For this purpose, a variant is proposed here which is essentially based on the application of neural networks, in particular convolutional networks (hereinafter also referred to as convolutional neural networks or “CNN” for short) from the field of machine learning and which are essentially used in the 3 and 4 is shown.

This variant is divided into a learning or training phase, in which characteristics typical for the process are first recognized or learned, and an inference phase, in which the spectra are analyzed based on the learned patterns during the actual measurement operation of the pressure measuring arrangement.

The basis for the training is the (high-pass filtered) Fourier coefficients of several chunks, each of which is composed of K frequency components (spectrograms) and is provided by the preprocessor. However, to train the network, in particular the convolutional network, several chunks must be aggregated into a sample. It can also be advantageous if a power spectrum is calculated. The FFT aggregator is set up to group N consecutive outputs of the preprocessor into two-dimensional spectrograms. Furthermore, a power for each frequency component can optionally be calculated for each output. After this step, the resulting two-dimensional spectrograms can be used to train the convolutional network. For training in the learning phase, it can initially be provided that the two-dimensional spectrograms are sent to an external computer via an unspecified data transfer, since the storage space and computing power of the microprocessor 162 may not be sufficient for training CNNs.

The two-dimensional spectrograms serve as inputs (samples) for the convolutional network. Each spectrogram is then assigned a binary label, where 0 stands for a free impulse line (negative class) and 1 for a blocked line (positive class). It is determined in advance at what level of blockage a sample belongs to the positive class. The spectrogram in the transition area between the two classes is assigned to the one from which the most Fourier coefficients come. In the event of a tie, the sample preferably belongs to the negative class. For subsequent training, the spectrograms (samples) can be mixed and randomly divided into a training, a validation and a test data set. The training data set can preferably contain 60% of the samples and the validation data set can contain a further 30% of the samples, so that the remaining 10% can form the test data set. The test dataset is required as an independent dataset because the validation data is used to adjust the hyperparameters, such as a batch size, and to detect under- and overfitting. Preferably, the spectrograms can be scaled in decibels when the dataset is generated. The logarithmic scale ensures better visual separation of small differences in performance, which simplifies training the networks.

5 shows an embodiment of a neural network that can be used both for training and in the inference phase. The network can preferably be a convolutional network. The network can receive as input, for example, a section of the spectrogram of dimension T × K × 1, where T describes the number of consecutive time steps (represented by different chunks) and K the number of Fourier coefficients remaining after high-pass filtering. The spectrogram section can then be processed by two convolution layers and reduced in dimension by max-pooling layers. After flattening, in which the multidimensional data structure is converted into a one-dimensional data structure, a three-stage classifier made up of fully connected layers of decreasing size can be used. This ultimately leads to a scalar output that is mapped into a value range (0; 1) by a sigmoid activation function. A limit for the assignment of a sample to the positive or negative class can then be set via a predeterminable threshold value η from this interval, so that a binary output can be generated that indicates whether a clogged differential pressure line is present or not.

For the convolution, classic convolutions (convolutional layers) are preferably not used. Instead, convolutional layers, especially so-called depthwise separable convolutions, are used. Depthwise Sepa rable convolutions are described, for example, in the publication “ MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications” by Andrew G. Howard, et al. or in the publication “ Rigid Motion Scattering for Texture Classification” by Laurent Slfre, Stéphane Mallat described. To do this, each of the N input channels is first transformed by a separate convolution with kernel size W × H × 1, where W and H denote the width and height of the kernels. The resulting results can then be combined to form a tensor of dimension T' × K' × N' before M convolutions with a kernel size 1 × 1 × N can be applied to the tensor. The result is ultimately a tensor of dimension T' × K' × M with M feature maps of size T' × K', as would have been the case when applying a classic (W × H × M) convolution. Both convolution layers can, for example, have kernels of size 3 × 3, with padding and stride each being one. In the first convolution layer, eight feature maps can be generated, which can be expanded to twelve by the second convolution. After each layer, a (2 × 2) max pooling can halve the width and height of the feature maps. In all convolution and fully connected layers, ReLu is preferably used as the activation function. This is excepted in the last layer, where the sigmoid function is used, as described above.

For the inference phase, in which the spectra are analyzed based on the learned patterns during the actual operation of the pressure measuring system, the trained model with the model parameters Θ, in particular the weights of the neural network, is transferred back to the computing unit of the pressure measuring system if the training was not carried out on the computing unit of the pressure measuring system. The model can then use the model parameters to directly classify the sample of the FFT aggregator at time t and generate an output via the threshold value η that indicates whether a clogged differential pressure line is present or not.

In principle, the output of the feature extractor, whether a clogged differential pressure line 25, 26 is present, can be sufficient for detecting a clogged differential pressure line 25, 26. However, this can produce incorrect results. For this reason, it can optionally be provided that in the classifier step following the feature extractor step, an additional smoothing of the output of the feature extractor is carried out using a mean value filter. The mean value filter aggregates the last M outputs that result from the feature extractor step and decides, depending on the threshold value, based on the mean value, whether a clogged differential pressure line is actually present or not.

list of reference symbols

1

pressure measuring arrangement

10

pressure transducer

11

sensor module

12, 13

process connection flanges

14

electronic module

16

two-wire line

18

process control system

20

differential pressure sensor

21

pipeline

22

aperture

23, 24

pressure tap channels

25, 26

differential pressure lines

110

pressure sensor element

112

measuring membrane

140

pressure transmission medium

150

detection unit

152

preprocessor

154

Feature Extractor

156

classifier

158

high-pass filter

160

operating and/or evaluation circuit

162

microprocessor

164

ADC

Bt-1, Bt

blocks or chunks

c1,t-1, cn,t

Frequency components of the frequency spectra of the blocks

CNN

convolutional neural network

Θ

model parameters

QUOTES INCLUDED IN THE DESCRIPTION

This list of 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 accepts no liability for any errors or omissions.

Cited patent literature

• EP 1 840 549 B1 [0002]
• DE 10 2013 110 059 A1 [0003]

Cited non-patent literature

• MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications” by Andrew G. Howard, [0036]
• Rigid Motion Scattering for Texture Classification” by Laurent Slfre, Stéphane Mallat [0036]

Claims (17)

A computer-implemented method for detecting a blocked differential pressure line of a pressure measuring arrangement, comprising the following steps: - detecting a pressure measuring signal by means of the pressure measuring arrangement (1), preferably for a defined period of time; - combining, preferably successive measured values of the pressure measuring signal into, preferably overlapping blocks (B _t-1 , B _t ), which preferably each have the same length; - generating, in particular, acoustic frequency spectra for each block, wherein each frequency spectrum has a plurality of frequency components (c _1,t-1 ; c _n,t ); - entering the, in particular acoustic frequency spectra of the blocks into a neural network which is configured, in particular trained, to detect a blocked differential pressure line based on the frequency spectra of the blocks; - generating an output which indicates whether or not a blocked differential pressure line is present. A computer-implemented method for training a neural network for detecting a blocked differential pressure line of a pressure measuring arrangement, comprising the following steps: - detecting a pressure measurement signal by means of the pressure measuring arrangement (1), preferably for a defined period of time; - combining, preferably successive measured values of the pressure measurement signal into, preferably overlapping blocks (B _t-1 , B _t ), which preferably each have the same length; - generating, in particular acoustic frequency spectra for each block, wherein each frequency spectrum has a plurality of frequency components (c _1,t-1 ; c _n,t ); - entering the, in particular acoustic frequency spectra of the blocks into a neural network for training the neural network based on the generated frequency spectra of the blocks. procedure according to claim 1 or 2 , wherein before the frequency spectra of the blocks are input into a neural network, the frequency spectra of several, preferably consecutive blocks are first aggregated to form a two-dimensional spectrogram, which has a time on the abscissa and an amplitude of the Fourier coefficients on the ordinate, and the spectrograms are used as input for the neural network. Method according to the preceding claim, wherein a power for each frequency component is calculated for the spectrograms. Method according to one or more of the preceding claims, wherein for training the neural network, each spectrogram is assigned a binary label which indicates whether a clogged differential pressure line is present or not. Method according to one or more of the preceding claims, wherein each spectrogram is processed by the neural network through several, preferably two, convolution layers and/or reduced in dimension by a max-pooling layer. Method according to one or more of the preceding claims, wherein the neural network is a convolutional network which is configured to use convolutional layers, in particular depth-separable convolutions. Method according to one of the preceding claims, wherein flattening is carried out in the convolutional network, in which a multi-dimensional output data structure of a respective convolutional layer of the convolutional network is converted into a one-dimensional output data structure. Method according to one or more of the preceding claims, wherein a multi-stage, preferably a three-stage classifier made up of fully connected layers of decreasing size generates a scalar output, in particular from the output data structure, which is mapped into a value range [0; 1], preferably by a sigmoid activation function, in order to generate the output which indicates whether a clogged differential pressure line is present or not. Method according to one or more of the preceding claims, wherein the frequency spectra are generated by means of a discrete Fourier transformation, in particular a fast Fourier transformation. Method according to one or more of the preceding claims, wherein the frequency spectra of the respective blocks (B _t-1 , B _t ) are high-pass filtered so that frequency components which are below a predefinable cut-off frequency are essentially filtered out or attenuated. Method according to the preceding claim, wherein in order to filter out the idle phases of the process for the blocks (B _t-1 , B _t ) at least one variance is calculated in each case, and a decision is made on the basis of the variance as to whether it is an idle phase of the process. Method according to the preceding claim, wherein the variance is determined with a configurable or adjustable threshold value in order to decide whether it is a rest phase of the process, wherein the threshold value is preferably in a range ≤ -70 dB, particularly preferably ≤ -80 dB, most preferably in the range from -80 to -90 dB. Data processing device with means for carrying out the method Method for detecting a clogged differential pressure line according to claim 1 and/or for carrying out the method for training a neural network according to claim 2 . Computer program with instructions which, when the program is executed by a computer, cause the computer to carry out the process for detecting a clogged differential pressure line according to claim 1 and/or the method for carrying out the method for training a neural network according to claim 2 to carry out. A computer-readable medium containing instructions that, when executed by a computer, cause the computer to perform the method for detecting a clogged differential pressure line according to claim 1 and/or the method for carrying out the method for training a neural network according to claim 2 to execute. Pressure measuring arrangement, at least comprising: - a pressure transducer (10) for detecting at least one media pressure of a medium; - an operating and/or evaluation circuit (14) for providing a pressure measurement signal which depends on the at least one media pressure; - at least one effective pressure line (25) which is connected to a pressure input (12) of the pressure transducer (10) in order to apply the at least one media pressure to the pressure transducer (10); - a detection unit (150) which is preferably designed as part of the operating and/or evaluation circuit and is set up to carry out the method for detecting a blocked effective pressure line according to claim 1 and/or the method for carrying out the method for training a neural network according to claim 2 to execute.

Images