Method for detecting blocked differential pressure line of pressure measuring device

Abstract

A computer-implemented method for detecting a blocked differential pressure line of a pressure measuring device, comprising the following steps: recording a pressure measurement signal by the pressure measuring device (1), preferably within a defined time period; -summarizing the measured values of the pressure measurement signal into preferably overlapping blocks (Bt-1, Bt), preferably in sequence, preferably with the same length; generating a spectrum, in particular an acoustic spectrum, for each block, wherein each spectrum has a plurality of frequency components (c1, t-1; cn, t); -inputting the spectrum of the block, in particular the acoustic spectrum, into a neural network configured to detect a blocked differential pressure line based on the spectrum of the block; generating an output specifying whether there is a blocked differential pressure line.

Description

Technical Field

[0001] The present invention relates to a computer-based method for detecting blockages in a differential pressure pipeline of a pressure measuring device, a data processing device having means for performing the method, a computer program, a computer-readable medium, and a pressure measuring device. Background Technology

[0002] European Patent EP 1 840 549 B1 protects an apparatus for recording blockages in differential pressure lines, wherein the apparatus comprises two pressure sensors for recording time series of static pressures in two differential pressure lines and a differential pressure sensor for recording time series of the difference between pressures in the two differential pressure lines; three calculation units for calculating the fluctuations of the two static pressures and the differential pressure based on the time series and the sum of squared fluctuations; a correlation coefficient calculation unit for determining the correlation coefficient between the time series of each static pressure and the differential pressure; and an evaluation unit for detecting blockages in the differential pressure lines based on the correlation and identifying which(s) of the differential pressure lines are blocked.

[0003] German patent application DE 10 2013 110 059 A1 describes a differential pressure measuring device that detects blocked differential pressure lines based on the correlation between a differential pressure signal—which represents the difference between the pressure of a first medium and the pressure of a second medium—and a temperature signal—which is related to the temperature of the differential pressure line.

[0004] Furthermore, other methods exist for detecting blockages in differential pressure lines. In particular, there are methods that operate based on changes in the frequency of pressure signal fluctuations in this case.

[0005] All methods essentially work toward a common goal; however, each has certain drawbacks in each case. Therefore, some methods are ineffective or only partially effective when the underlying process is too dynamic and subject to large fluctuations. Other methods only work within a specific pressure signal range, which dynamic processes typically exceed. Similarly, other methods suffer from drawbacks when generating resonance through long differential pressure lines, affecting signal power and sometimes potentially artificially increasing it. Summary of the Invention

[0006] The purpose of this invention is to solve the above-mentioned problems.

[0007] According to the present invention, this objective is achieved by the method for detecting a blocked differential pressure pipeline as defined in claim 1, the method for training a neural network as defined in claim 2, the data processing device as defined in claim 14, the computer program as defined in claim 15, the computer-readable medium as defined in claim 16, and the pressure measuring device as defined in claim 17.

[0008] The computer-based method for detecting blockages in a differential pressure pipeline of a pressure measuring device according to the present invention includes the following steps:

[0009] - The pressure measurement signal is recorded by means of a pressure measuring device, preferably within a limited time period;

[0010] - The measured values ​​of the pressure measurement signal, preferably consecutively following the measured values, are summarized into preferably overlapping blocks, which preferably have the same length;

[0011] - Generate a spectrum for each block, especially an acoustic spectrum, where each spectrum has multiple frequency components;

[0012] - The spectrum of the block, especially the acoustic spectrum, is input into a neural network, which is configured, and in particular trained, to detect blocked differential pressure lines based on the spectrum of the block.

[0013] - Generates an output that specifies whether there is a blockage in the differential pressure line.

[0014] The present invention also relates to a computer-based method for training a neural network to detect blockages in a differential pressure pipeline of a pressure measuring device, comprising the following steps:

[0015] - The pressure measurement signal is recorded by means of a pressure measuring device, preferably within a limited time period;

[0016] - The measured values ​​of the pressure measurement signal, preferably consecutively following the measured values, are summarized into preferably overlapping blocks, which preferably have the same length;

[0017] - Generate a spectrum for each block, especially an acoustic spectrum, where each spectrum has multiple frequency components;

[0018] - The spectrum of the block, especially the acoustic spectrum, is input into the neural network to train the neural network based on the spectrum of the resulting block.

[0019] According to the present invention, a method is provided for recording and evaluating the spectrum of a pressure measurement signal from a pressure measuring device. This spectrum is preferably an acoustic spectrum, located in the region less than 20 kHz, particularly less than 2.5 kHz, especially in the 0-2 kHz range, and very particularly in the 5 Hz to 200 Hz range. This (acoustic) spectrum characterizes the pressure measuring device and essentially represents the fingerprint of the pressure measuring transducer. Changes in the pressure measuring device affect the spectrum, particularly the acoustic spectrum. According to the present invention, this spectrum is used to train a neural network to detect blocked differential pressure lines.

[0020] An advantageous form of the method of the present invention is provided in which, before inputting the spectrum of a block into a neural network, the spectra of multiple blocks, especially consecutively following blocks, are first aggregated into a two-dimensional spectrogram having time on the horizontal axis and magnitude of Fourier coefficients on the vertical axis, and the spectrogram is used as the input to the neural network.

[0021] In particular, the form of this embodiment can provide power calculation for each frequency component of the spectrum.

[0022] Another advantageous form of the method of the present invention provides that, in order to train the neural network, each spectrogram is associated with a binary label that indicates whether there is a blockage in the differential pressure line.

[0023] Another advantageous form of the method of the present invention is provided, wherein each spectrogram is processed by a neural network through multiple convolutional layers, preferably two convolutional layers, and / or dimensionality reduced by a max pooling layer.

[0024] Another advantageous form of the method of the present invention provides that the neural network is a convolutional network, which is adapted to utilize convolutional layers, especially depthwise separable convolutions. In particular, this embodiment can provide that flattening is performed in the convolutional network, in which case the multidimensional output data structure of the convolutional layers of the convolutional network is transformed into a one-dimensional output data structure.

[0025] Another advantageous form of the method of the present invention provides that, by reducing the size of a multi-stage, preferably three-stage, fully connected layer classifier, in particular from the output data structure, a scalar output is generated, which is preferably mapped to the [0;1] value range by a sigmoid activation function, in order to generate an output specifying whether a differential pressure line is blocked.

[0026] Another advantageous form of the embodiment of the method of the present invention provides that the spectrum is generated by means of the discrete Fourier transform, especially the fast Fourier transform.

[0027] Another advantageous form of the method of the invention provides that the spectrum of the block is high-pass filtered such that frequency portions below a predetermined threshold frequency are substantially filtered out, or at least attenuated. In particular, this form of the embodiment may provide that, in order to filter out quiescent phases of the process, at least a variance is calculated in each case, and a determination is made based on the variance as to whether this is a quiescent phase of the process. In particular, this form of the embodiment may provide that the variance is compared with a configurable or selectable threshold to determine whether a quiescent phase of the process is in, wherein the threshold is preferably located in the region ≤-70 dB, particularly preferably in the region ≤-80 dB, and very particularly preferably in the range of -80 to -90 dB.

[0028] The present invention also relates to a data processing apparatus having means for performing a method for detecting a blocked differential pressure pipeline according to one of the above embodiments and / or for performing a method for training a neural network.

[0029] The present invention also relates to a computer program having instructions that, when executed by a computer, cause the computer to perform a method for detecting a blocked differential pressure pipeline according to one of the above embodiments and / or a method for training a neural network.

[0030] The present invention also relates to a computer-readable medium containing instructions that, when executed by a computer, cause the computer to perform a method for detecting a blocked differential pressure pipeline according to one of the above embodiments and / or a method for training a neural network.

[0031] The present invention also relates to a pressure measuring device, comprising:

[0032] - A pressure measuring transducer used to record at least one medium pressure;

[0033] - Operation and / or evaluation circuitry for providing a pressure measurement signal that depends on the pressure of at least one medium;

[0034] - At least one differential pressure line connected to the pressure input of a pressure measuring transducer to supply at least one medium pressure to the pressure measuring transducer;

[0035] - A detection unit, which is preferably formed as part of an operating and / or evaluation circuit, and is adapted to perform a method for detecting a blocked differential pressure line according to one of the above embodiments and / or a method for training a neural network. Attached Figure Description

[0036] The invention will now be explained in more detail with reference to the accompanying drawings, which are shown below:

[0037] Figure 1 This is an overall view of an example embodiment of the pressure measuring device of the present invention.

[0038] Figure 2 Details of the pressure measuring device in the form of a sensor module and an electronic module electrically connected to the sensor module.

[0039] Figure 3 This is a general block diagram of a computer-based method for detecting blockages in differential pressure pipelines used in pressure measuring devices.

[0040] Figure 4 This is another block diagram of a method for detecting blockages in differential pressure lines of pressure measuring devices, which illustrates the learning phase and the inference phase.

[0041] Figure 5 This is another block diagram of a method for detecting blockages in differential pressure lines of pressure measuring devices using neural networks.

[0042] Figure 6 It is a spectrum of the average block of cleaned-up data transformed into the frequency domain by means of discrete Fourier transform. Detailed Implementation

[0043] Figure 1 A greatly simplified overall schematic diagram of the pressure measuring device is shown. Figure 1 The image shows a differential pressure measuring device. Although the following description is based on a differential pressure measuring device, the invention is not limited thereto, but rather, equally applicable to any other pressure measuring device, especially absolute pressure and / or relative pressure measuring devices. Figure 1 The differential pressure measuring device shown includes a differential pressure measuring transmitter / transducer 10 with a sensor module 11 disposed 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 housed within the housing of the differential pressure measuring transducer. The electronic module can be connected to a process control system 18 via a two-wire line 16, through which the electronic module 14 can communicate and be powered. This two-wire line can operate as a fieldbus according to the Profibus or Foundation fieldbus standard or according to the HART standard. Such differential pressure measuring transducers are known in themselves and are manufactured and sold by the applicant, for example, under the brand Deltabar.

[0044] The differential pressure measuring device may also include a differential pressure generator 20 installed in the conduit 21. The differential pressure generator includes a diaphragm 22, a first pressure sensing tube 23 on the high-pressure side of the diaphragm 22, and a second pressure sensing tube 24 on the low-pressure side of the diaphragm 22. A high-pressure side process connection flange 12 is connected to the high-pressure side pressure sensing tube 23 via a high-pressure side differential pressure line 25, and a low-pressure side process connection flange is connected to the low-pressure side pressure sensing tube 24 via a low-pressure side differential pressure line 26. The concepts of "high-pressure side" and "low-pressure side" relate to the differential pressure generated by the fluid (from left to right in the figure), which is proportional to the square of the flow velocity and is on the order of, for example, about 10 to 100 mbar. The static pressure superimposed on this flow-related differential pressure can be on the order of, for example, from 1 bar to several hundred bar.

[0045] Figure 2 The sensor module 11 and the electronic module 14 electrically connected to the sensor module 11 are shown in more detail. For the sake of simplicity, Figure 2 The two modules shown are arranged adjacent to each other, and therefore... Figure 1The difference is as follows. In this example, the sensor module uses a piezoresistive pressure sensor element 110 with a measuring diaphragm 112, wherein the pressure sensor element 110 has a resistive element in a bridge circuit to convert the pressure-related deformation or deflection of the measuring diaphragm into an electrical sensor module signal. Instead of a piezoresistive pressure sensor, a capacitive pressure sensor can also be provided, in which case the measuring diaphragm includes electrodes, the capacitance of which relative to the electrodes on the bending platform is a measure of the pressure-related deformation or deflection of the measuring diaphragm. Details of the electrical transducer are known to those skilled in the art of pressure measurement and need not be explained further here. The sensor module outputs a sensor module signal based on the recorded differential pressure to an electronic module 14, wherein the operation and / or evaluation circuit 160 of the electronic module 14 generates a differential pressure measurement signal representing the differential pressure based on the sensor module signal and can output it, for example, via a two-wire line 16 to a process control system 18. For this purpose, the operation and / or evaluation circuit may have a computing unit, for example, in the form of a microprocessor 162, for processing the sensor module signal of the pressure sensor 110 digitized by means of an ADC 164.

[0046] The pressure measuring device includes an additional detection unit 150, which is adapted to perform the computer-implemented method described below. The detection unit 150 may be, for example... Figure 2 As shown, it is part of the operation and / or evaluation circuitry 160, or alternatively implemented as a separate unit (e.g., in the cloud).

[0047] The computer implementation method includes several method steps, including: a preprocessing step performed in preprocessor 152 (see...). Figure 3 In the case where the pressure measurement signal or differential pressure measurement signal is converted into a spectrum; the feature extraction step performed in feature extractor 154 (see...) Figure 4 In the case where features are derived from the spectrum using a neural network, an optional classifier step can also be performed in classifier 156 to determine whether differential pressure lines 25 and 26 are blocked based on the features derived above. In this case, the method steps are executed sequentially. Figure 3 A schematic diagram of a computer implementation method for detecting blockages in differential pressure lines 25 and 26 of pressure measuring device 1 is shown.

[0048] In the first step, the pressure measurement signal is sampled at the highest possible sampling rate to obtain continuous measurement values, preferably digital measurement values, and especially preferably measurement values ​​digitized by an ADC164. In this case, the sampling rate is preferably selected such that its frequency is at least greater than 10 Hz, preferably at least 50 Hz, and especially preferably at least 100 Hz.

[0049] To prevent stagnant phases in the process, the pressure values ​​will be determined by pressure measuring device 1, making it difficult to detect blockages in differential pressure lines 25 and 26. These stagnant phases must be detected and filtered out as quickly as possible. This can occur in various ways. A simple method for this purpose is to combine the values ​​preferably measured continuously in the preprocessing step into preferably overlapping blocks B. t-1 B t (Hereinafter also referred to as chunks), which preferably have equal lengths. In this case, chunk B t-1 B t Forming indivisible units, and block B t-1 B t All measurements will be processed together. Then, in the preprocessor step, for each data block B... t-1 B t Calculate the variance, and optionally the mean. Then compare the variance to a configurable or optional threshold. If this is the case, block B... t-1 Or B t If the variance is below the threshold, then block B is ignored. t-1 Or B t This is because it has a high probability of being in a static phase of the process. In this case, the threshold may vary depending on the process and from process to process. For example, the threshold for a fill level process, where the fill level is determined by means of a pressure measuring device, may differ from the threshold for a flow process, where the flow rate is determined by means of a differential pressure measuring device. Typically, however, the threshold is in the range of ≤-70 dB, preferably in the range of ≤-80 dB, and particularly preferably in the range of -80 to -90 dB.

[0050] Remaining block B t-1 B t The mean is adjusted in the preprocessor step, and the result is transformed to the frequency domain using a discrete Fourier transform (e.g., a fast Fourier transform (FFT)) to obtain each block B required for subsequent feature extraction steps. t-1 B t Fourier coefficients. Figure 6 Such a spectrum is illustrated in the example. Following the Fourier transform, in the preprocessing step, a high-pass filter 158 can be optionally applied, operating at a preset threshold frequency to filter out the actual pressure values. For example, the preset threshold frequency can be selected such that it lies in the range of a few Hz, preferably less than 5 Hz, and particularly preferably in the range of 1-5 Hz. The output after preprocessing consists of the Fourier coefficients of the blocks (in certain cases, after high-pass filtering).

[0051] In the feature extraction step following the preprocessing step, the features required to detect blockages in the differential pressure lines 25 and 26 are identified. To this end, a variant is provided here, primarily based on neural networks in the field of machine learning, especially convolutional networks (hereinafter also referred to as convolutional neural networks or simply "CNNs"), and its main features are as follows: Figure 4 and Figure 5 As shown.

[0052] This variant is divided into a learning or training phase, in which typical characteristics of the process are first detected and learned, and an inference phase, in which the spectrum is analyzed based on the learned patterns during actual measurement operations of the pressure measuring device.

[0053] Training is based on (high-pass filtered) Fourier coefficients in multiple blocks, each consisting of K frequency components (spectral maps) provided by a preprocessor. To train the network, especially the convolutional network, these blocks must be aggregated into a single sample. Furthermore, calculating the power spectrum may be advantageous. An FFT aggregator is suitable for grouping N consecutive follow-up outputs from the preprocessor into two-dimensional spectrograms. Additionally, optionally, power can be calculated for each frequency component for each output. Following this step, the resulting two-dimensional spectrogram can be used to train the convolutional network. During the training phase of the learning process, it is advisable to first transmit the two-dimensional spectrogram to an external computer via a suitable data transfer method, as the storage space and computing power of the microprocessor 162 are insufficient for training a CNN in the given circumstances.

[0054] Two-dimensional spectrograms are used as input (samples) to the convolutional network. Each spectrogram is then assigned a binary label, where 0 represents a free-pulse line (negative class) and 1 represents a blocked line (positive class). The degree of blocking required for a sample to belong to the positive class is determined earlier. Spectrograms located in the transition region between the two classes are assigned to the class from which the most Fourier coefficients originate. In the event of a tie, the sample is preferably classified as negative. During subsequent training, the spectrograms (samples) can be mixed and randomly distributed across training, validation, and test datasets. The training dataset can preferably contain 60% of the samples, and the validation dataset contains another 30%, leaving the remaining 10% to form the test dataset. The test dataset needs to be a separate dataset because the validation data is used to fit hyperparameters (such as batch size) and to detect underfitting and overfitting. Preferably, the spectrograms are scaled in decibels when generating the dataset. Logarithmic scaling provides better differentiation for small power differences, thus benefiting network training.

[0055] Figure 5An embodiment of a neural network is shown, which can be used for both training and inference phases. The network can preferably be a convolutional network. For example, the network can use a spectral image segment of dimension T×K×1 as input, where T is the number of consecutive time steps (mapped by different blocks), and K is the number of Fourier coefficients remaining after high-pass filtering. The spectral image segment is then processed by two convolutional layers and dimensionality reduced by a max-pooling layer. After flattening the multidimensional data structure into a one-dimensional data structure, a three-level classifier with fully connected layers of decreasing size can be followed. This ultimately results in a scalar output in the form of a sigmoid activation function in the range (0;1). This is achieved via a predetermined threshold from the interval. Then a constraint can be set to assign the sample to the positive or negative class, which can produce a binary output that specifies whether there is a blocked differential pressure line.

[0056] Classical convolutions are preferably not used for convolution. Instead, preferred convolutional layers are utilized, especially so-called depthwise separable convolutions. Depthwise separable convolutions are described, for example, in the publication "MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications" by Andrew G. Howard et al., or by Laurent KIfre and... The publication is titled "Rigid Motion Scattering for Texture Classification". To this end, each of the N input channels is first transformed by a separable convolution with a kernel size of W×H×1, where W and H refer to the width and height of the kernel. The results are then summed together into a convolution with dimension W... We first construct a tensor, and then apply M convolutions with kernel size 1×1×N to this tensor. The final result is a tensor with dimension M. A tensor, having M elements of size M. The feature map is the same as that of the case using classic (W×H×M) convolution.

[0057] For example, two convolutional layers can have kernels of size 3×3, with padding and stride of 1 in each case. In the first convolutional layer, 8 feature maps can be generated, which can be expanded to 12 by the second convolution. After each layer, a (2×2) max-pooling layer can halve the width and height of the feature maps. ReLU is preferably used as the activation function in all convolutional and fully connected layers, except in the case of the last layer, where a sigmoid function is used as described above.

[0058] For the inference phase, during the actual operation of the pressure measuring device, the spectrum is analyzed based on the learned model. When the computational unit of the pressure measuring device has not been trained, it will have model parameters. In particular, the trained model of the neural network weights is transmitted back to the computational unit of the pressure measurement device. Then, the model with the model parameters can directly classify the samples from the FFT aggregator at time point t, and pass a threshold... It generates an output that specifies whether there is a blockage in the differential pressure line.

[0059] Basically, the output of the feature extractor, indicating whether there is a blockage in the differential pressure lines 25 and 26, is sufficient to detect blockages in these lines. However, erroneous results may occur. For this reason, alternatively, in the classifier step following the feature extractor step, the output of the feature extractor can be further smoothed using an averaging filter. The averaging filter aggregates the last M outputs from the feature extractor step and determines the threshold based on the average value, regardless of whether a blockage in the differential pressure line actually exists.

[0060] List of reference symbols

[0061] 1. Pressure measuring device

[0062] 10 Pressure Measurement Transducer

[0063] 11 sensor modules

[0064] 12, 13 Process connection flanges

[0065] 14 electronic modules

[0066] 16 double conductor circuit

[0067] 18 Process Control Systems

[0068] 20 Differential Pressure Generator

[0069] 21 pipelines

[0070] 22 membranes

[0071] 23, 24 pressure sensor tubes

[0072] 25, 26 Differential pressure lines

[0073] 110 pressure sensor element

[0074] 112 Measuring Diaphragm

[0075] 140 pressure transmission medium

[0076] 150 detection units

[0077] 152 preprocessor

[0078] 154 Feature Extractor

[0079] 156 classifier

[0080] 158 high-pass filter

[0081] 160 Operation and / or Evaluation Circuits

[0082] 162 microprocessor

[0083] 164ADC

[0084] B t-1 B t Blocks or groups

[0085] c 1,t-1 ;c n,t Frequency components of the block's spectrum

[0086] CNN (Convolutional Neural Network)

[0087] Model parameters

Claims

1. A computer-implemented method for detecting a plugged differential pressure line of a pressure measuring device, comprising the following steps: - recording pressure measuring signals by means of the pressure measuring device (1), preferably over a defined time period; - aggregating the measured values, preferably successive measured values, of the pressure measurement signal to preferably overlapping blocks (B t-1 , B t ), which preferably have the same length; - generating a spectrum, in particular an acoustic spectrum, for each block, wherein each spectrum has a plurality of frequency components (c 1,t-1 ; c n,t ) ; - inputting the spectrum, in particular the acoustic spectrum, of the block into a neural network, the neural network being configured, in particular trained, to detect a plugged differential pressure line based on the spectrum of the block; - generating an output specifying whether a plugged differential pressure line is present or not.

2. A computer-implemented method for training a neural network to detect a plugged differential pressure line of a pressure measuring device, comprising the following steps: - recording pressure measuring signals by means of the pressure measuring device (1), preferably over a defined time period; - aggregating the measured values, preferably successive measured values, of the pressure measurement signal to preferably overlapping blocks (B t-1 , B t ), which preferably have the same length; - generating a spectrum, in particular an acoustic spectrum, for each block, wherein each spectrum has a plurality of frequency components (c 1,t-1 ; c n,t ) - inputting the spectrum, in particular the acoustic spectrum, of the block into a neural network for training the neural network based on the generated spectrum of the block.

3. The method of claim 1 or 2, wherein, Before inputting the spectrum of the block into the neural network, a plurality of blocks, in particular the spectra of consecutively following blocks, are first aggregated into a two-dimensional spectrum plot having the time on the abscissa and the amplitude of the Fourier coefficients on the ordinate, and this spectrum plot is used as input to the neural network.

4. The method according to the preceding claim, wherein, The power of each frequency component is calculated for the spectrum plot.

5. The method according to one or more of the preceding claims, wherein, For training the neural network, each spectrum plot is associated with a binary label specifying whether a plugged differential pressure line is present or not.

6. The method according to one or more of the preceding claims, wherein, Each spectrum plot is processed by the neural network by a plurality of convolutional layers, preferably two convolutional layers, and / or is reduced in dimension by a max-pooling layer.

7. The method according to one or more of the preceding claims, wherein, The neural network is a convolutional network adapted to utilize convolutional layers, in particular depthwise separable convolutions.

8. The method according to one of the preceding claims, wherein, In the convolutional network, a flattening is performed, in which case the multi-dimensional output data structure of the convolutional layers of the convolutional network is converted into a one-dimensional output data structure.

9. The method according to one or more of the preceding claims, wherein, From the output data structure, a scalar output is generated by a multi-stage, preferably three-stage, fully connected layer classifier, which maps preferably by a sigmoid activation function to a [0; 1] value range in order to generate an output specifying whether a plugged differential pressure line is present or not.

10. The method according to one or more of the preceding claims, wherein, The spectrum is generated by means of a discrete Fourier transform, in particular a fast Fourier transform.

11. The method according to one or more of the preceding claims, wherein, The spectrum of the block (B t-1 , B t ) is high-pass filtered such that the frequency portion below a predetermined limit frequency is substantially filtered out or at least attenuated.

12. The method according to the preceding claim, wherein, To filter out the stationary phase of the process of the blocks (B t-1 , B t ) at least a variance is calculated in each case and based on the variance it is decided whether this is a stationary phase of the process.

13. The method according to the preceding claim, wherein, The variance is compared to a configurable or selectable threshold value in order to decide whether a stationary phase of the process is present, wherein the threshold value is preferably located in the region of < -70 dB, in particular preferably in the region of < -80 dB, very particularly preferably in the range of -80 to -90 dB.

14. A data processing device having means for carrying out the method for detecting a plugged differential pressure line according to claim 1 and / or for carrying out the method for training a neural network according to claim 2.

15. A computer program having instructions which, when the program is executed by a computer, cause the computer to carry out the method for detecting a plugged differential pressure line according to claim 1 and / or the method for training a neural network according to claim 2.

16. A computer-readable medium containing instructions which, when executed by a computer, cause the computer to carry out the method for detecting a blocked differential pressure line according to claim 1 and / or the method for training a neural network according to claim 2.

17. A pressure measuring device, comprising: - a pressure measuring transducer (10) for recording at least one medium pressure of a medium; - an operating and / or evaluation circuit (14) for providing a pressure measuring signal dependent on the at least one medium pressure; - at least one differential pressure line (25) connected to a pressure input (12) of the pressure measuring transducer (10) in order to supply the at least one medium pressure to the pressure measuring transducer (10); - a detection unit (150), which is preferably formed as part of the operating and / or evaluation circuit and is adapted to carry out the method for detecting a blocked differential pressure line according to claim 1 and / or the method for training a neural network according to claim 2.

Images