Method for monitoring the condition of a sensor

Abstract

A method for monitoring the state of a measuring instrument using a state parameter which has cross-sensitivities to cross-sensitivity parameters comprises: Install 210 of the measuring device at a measuring point in a process plant; Capture 220 from an initial data set; which includes state parameter measurements and cross-sensitivity parameter measurements of multiple cross-sensitivity parameters. Dividing the initial data set into a training set, a validation set, and a test set; Conducting a model training session of 240 using the training set; Determine and store 250 of an empirical probability distribution function (EPDF) of prediction errors between a state parameter prediction value predicted by the trained model based on cross-sensitivity parameter measurements and a state parameter measurement from the test set; Acquisition of monitoring data, each comprising cross-sensitivity parameter measurements and an associated condition parameter measurement; Determining a state parameter prediction value based on the cross-sensitivity parameter measurements; Determining the prediction error between the state parameter prediction value and the corresponding state parameter measurement value; and Determine 260 a probability value using the probability distribution function for the prediction error.

Description

[0001] The present invention relates to a method for monitoring the condition of a sensor installed at a measuring point. US 2007 / 028663 A1 and DE 602 18 050 T2 disclose methods for monitoring the condition of a sensor in which a media density determined by the measuring device is compared with a reference value. A generic method is disclosed, for example, in WO 2012 / 062551 A1 and EP 3 894 802 B1. According to the latter document, the oscillation amplitude of the oscillator at a frequency that deviates from a resonant frequency of a bending vibration mode by a factor of , is a suitable parameter for condition monitoring.The method according to EP 3 894 802 B1 for monitoring the condition of a sensor installed at a measuring point comprises identifying tuples that contain a value of at least one physical parameter with which cross-sensitivities of the condition parameter exist and an associated value of the condition parameter, as well as assigning the tuples to a value range of the at least one physical parameter and forming a reference value of the condition parameter for this value range of the at least one physical parameter using the value of the condition parameter of the tuple if no valid reference value of the condition parameter exists for this value range; or comparing the value of the condition parameter of the tuple with the reference value if a reference value of the condition parameter exists for this value range; and generating a finding depending on the result of the comparison.The finding can, for example, be the identification of a fault condition if the comparison reveals a significant deviation from the reference value. This prior art method has limitations insofar as, firstly, the definition of the value ranges is arbitrary, and, particularly in the case of higher-dimensional cross-sensitivities (i.e., cross-sensitivities to several physical parameters), this leads to a large number of cells with reference values ​​in the reference space. Furthermore, the values ​​of the state parameters in the individual cells are isolated snapshots, the temporal fluctuations of which are not necessarily uniform across the entire state space. Thus, there is a risk of false positives when identifying fault conditions. The object of the invention is to remedy this.

[0002] The problem is solved according to the invention by the method according to independent claim 1.

[0003] The inventive method for monitoring the state of a measuring device based on a state parameter which exhibits cross-sensitivities to cross-sensitivity parameters comprises: installing the measuring device at a measuring point in a process plant; acquiring an initial data set which includes state parameter measurements and cross-sensitivity parameter measurements of several cross-sensitivity parameters; dividing the initial data set into a training set, a validation set, and a test set; performing model training; determining and storing an empirical probability distribution function (EPDF) of prediction errors between a state parameter prediction value predicted by the trained model based on cross-sensitivity parameter measurements and a state parameter measurement value from the test set;Acquisition of monitoring data, each comprising cross-sensitivity parameter measurements and a corresponding state parameter measurement; determination of a state parameter prediction value based on the cross-sensitivity parameter measurements; determination of a prediction error between the state parameter prediction value and the corresponding state parameter measurement; and determination of a probability value using the probability distribution function for the prediction error.

[0004] In a further development of the invention, the method comprises comparing the probability with a threshold value, whereby falling below the threshold value is considered an indication of a change of state.

[0005] In a further development of the invention, a number of consecutive instances of falling below the threshold value is recorded, and a change of state is signaled when the number exceeds a limit value.

[0006] In a further development of the invention, signaling includes the request for a maintenance measure.

[0007] In a further development of the invention, a forecast for the future course of the prediction errors or their probability is created based on the temporal development of the prediction errors.

[0008] In a further development of the invention, a time-based forecast for maintenance requirements is created and signaled based on the prediction.

[0009] In a further development of the invention, the measuring device comprises a Coriolis mass flow meter with at least one oscillating measuring tube for guiding a medium, wherein the state parameter Z(t) is indicative of the measuring tube stiffness of the oscillating measuring tube.

[0010] In a further development of the invention, the transverse sensitivity parameters Cp include i , several of the following parameters include: temperature of at least one measuring tube, temperature gradients, pressure, density and viscosity of the medium, as well as carrier body temperature, temperature of electronic components of the measuring and operating circuit, vibration damping and sensor asymmetry.

[0011] In a further development of the invention, the cross-sensitivity parameters Cp are i , can be detected using components of the Coriolis mass flow meter itself or by means of additional sensors.

[0012] In a further development of the invention, the maintenance requirement includes a recalibration of the Coriolis mass flow meter.

[0013] As an example of a state that can be monitored with a state parameter Z(t), the so-called HBSI (High-Frequency Stability Index) serves, as already mentioned. This essentially describes a measurement of the pipe stiffness in a Coriolis measuring device. This pipe stiffness is directly and almost linearly correlated with the calibration factor of the measuring device. However, the sensitivity of this relationship (slope) depends on several factors, such as the sensor type, the diameter, but also on the process conditions and the wear mode, i.e., abrasion, corrosion, and deposits, which affect the pipe stiffness and thus the k-factor of the measuring device. Apart from investigating the degradation modes and their dynamics, the calibration factor, or HBSI, has a direct influence on the metrological performance of the measuring device.This means that an HBSI drift implies a calibration factor drift, i.e., a deterioration in the meter's measurement performance to the extent that the meter may no longer perform as expected, which is generally expressed as MPE (Maximum Permissible Error) or measurement uncertainty.

[0014] According to the invention, a model is developed using machine learning algorithms, which makes it possible to detect in real time when the measurement performance may be impaired compared to the initial state of the measuring instrument.

[0015] For this purpose, an ML model is trained to determine state parameter prediction values ​​based on cross-sensitivity parameter values, for example HBSI prediction values, and to compare these with state parameter measurements, i.e. HBSI measurements.

[0016] The input parameters are all influencing factors or cross-sensitivity parameters except for the condition parameter (HBSI) itself. This is the main difference between physical HBSI measurement and HBSI prediction. If the integrity of the pipe is compromised during operation, the physical HBSI reflects the change in pipe stiffness, while the soft HBSI remains unchanged for the given process conditions. Therefore, the error between the soft HBSI and the physical HBSI will increase if the measurement performance actually deteriorates.

[0017] As with any machine learning algorithm, the model must be trained on actual process conditions that can occur during a period in which the measuring instrument has not yet experienced any degradation in its metrological performance. This period should be relatively short, e.g., a few weeks or months, or at least without any significant change in the calibration factor. This is a necessary prerequisite for the application of the method.

[0018] The invention will now be explained with reference to the exemplary embodiments shown in the drawings. These show: Fig. 1. An embodiment of a measuring point for carrying out the method according to the invention; Fig. 2a: A diagram with state parameter measurements Z(t) (curve a) and cross-sensitivity parameter measurements cp i (t) for two different cross-sensitivity parameters CP i (curves b and c); Fig. 2b: A bar chart showing weights W(CP) i ) shows different cross-sensitivity parameters in the trained model; Fig. 2c: A diagram showing state parameter measurements Z(t) (curve a) and state parameter prediction values ​​Z(t) (curve b) after a successful training phase; Fig. 3a: A diagram with state parameter measurements Z(t) (curve a) and state parameter prediction values ​​Z(t) (curve b) and the associated probability (curve c) according to EPDF for a measuring device in a stable state; Fig. 3b: A diagram showing state parameter measurements Z(t) (curve a) and state parameter prediction values ​​Z(t) (curve b) and the associated probability (curve c) according to EPDF for a measuring device in a deteriorating state; and Fig. 4: A flowchart of an embodiment of the method according to the invention.

[0019] The in Fig. Figure 1 shows an example of a measuring point 1 for carrying out the method according to the invention, which is arranged in a pipeline 10 in which a medium flows. The measuring point 1 comprises a Coriolis mass flow sensor 20, which is configured to detect not only mass flow rates but also the density of a medium. Such a Coriolis mass flow sensor 20 is manufactured, for example, by the applicant under the designation Promass F, Promass Q, or Promass X. The Coriolis mass flow sensor 20 comprises at least one oscillator which includes two curved, parallel measuring tubes 22 in a housing 24, which can be excited to bending vibrations. The housing comprises a rigid support body 25 in which the measuring tubes 22 are mounted on the inlet and outlet sides. The mass flow rate, density, and viscosity of the medium can be determined from the vibration behavior of the measuring tubes 22 in a manner known per se.The flow sensor 20 is shown in the drawing with a horizontal flow direction and a downward-pointing measuring tube bend. Of course, the measuring tube bend can also point upwards for improved drainage. Likewise, the flow sensor can also be arranged with a vertical flow direction. The measuring point 1 optionally includes two pressure transmitters 32 and 34, with the Coriolis mass flow sensor 22 positioned between the pressure transmitters. The pressure of the medium in the Coriolis mass flow sensor is essentially the average of the measured values ​​of the two pressure transmitters.Since the Coriolis mass flow sensor has a constricted flow cross-section compared to the pipeline, it forms a restrictor between the two pressure transmitters 32, 34. The viscosity of the medium can then be determined from the difference between their pressure readings, as the flow velocity is also known via the mass flow rate and the density of the medium. The pressure sensors of the pressure transmitters can be absolute or relative pressure sensors. Instead of the two pressure transmitters, a single pressure transmitter can also be used, for example, an inlet-side pressure transmitter 32. In this case, the pressure prevailing in the measuring tubes can be determined based on a pressure reading and other medium parameters, as disclosed, for example, in German patent application DE 10 2010 000 759 A1.The pressure readings from one or both pressure transmitters are transmitted to the Coriolis mass flow sensor and / or to a higher-level unit 240.

[0020] The Coriolis mass flow sensor 20 can further include temperature sensors for recording temperature measurements, for example an inlet-side measuring tube temperature sensor T22a on one of the measuring tubes 22 and an outlet-side measuring tube temperature sensor T22b on one of the measuring tubes 22, a carrier body temperature sensor T25 on the carrier body 25, and an electronic temperature sensor T26 on a measuring and operating circuit 26.

[0021] The measuring tube temperature sensors allow for the estimation of the measuring tube temperature, the medium temperature, and any temperature gradients along the measuring tubes. The support body temperature sensor also enables the detection of temperature gradients between measuring tubes 22 and the support body 25, which can result in mechanical stresses.

[0022] The Coriolis mass flow sensor 20 further comprises the measuring and operating circuit 26, which is configured at least to operate the flow sensor 20, to determine measured values ​​for the mass flow and, if applicable, the density, and to output the determined measured values ​​to the higher-level unit 40. Furthermore, the measuring and operating circuit 26, or optionally in combination with the higher-level unit 40, is configured to carry out the method according to the invention.

[0023] Diagram in Fig. Figure 2a shows state parameter measurements Z(t) (curve a) and cross-sensitivity parameter measurements cp i (t) for two different cross-sensitivity parameters CP i (Curves b and c). The state parameter can be measured, in particular, as the vibration signal amplitude of the measuring tube vibration at a frequency that corresponds to 1.15 times the resonance frequency of the so-called bending vibration mode. The cross-sensitivity parameters here can be, for example, the two different temperature values, such as those of the measuring tube and the support body.

[0024] The bar chart in Fig. 2b shows which weights W(CP) i The trained model exhibits various cross-sensitivity parameters. Some temperature values ​​are prominent, while others are negligible. These latter values ​​no longer need to be recorded after the training phase is complete.

[0025] The diagram in Fig. Figure 2c shows a good agreement between state parameter measurements Z(t) (curve a) and state parameter prediction values ​​Z(t) (curve b) based on the trained model after a successful training phase.

[0026] The diagram in Fig. Figure 3a also shows good agreement between state parameter measurements Z(t) (curve a) and state parameter prediction values ​​Z(t) (curve b) based on the trained model. Accordingly, the associated probability (curve c) according to EPDF is high, which accurately describes the situation of a measuring device in a stable state.

[0027] The diagram in Fig. In contrast, 3b shows increasing deviations between measured state parameter values ​​Z(t) (curve a) and predicted state parameter values ​​Z(t) (curve b). Accordingly, the associated probability (curve c) decreases according to EPDF, thus accurately describing the situation of a measuring device in a deteriorating state.

[0028] An embodiment 200 of the method according to the invention will now be described using the flowchart in Fig. 4 described.

[0029] First, the measuring device is installed at a measuring point in a process plant. The measuring device's performance is known thanks to recent calibration and, if necessary, adjustment. Next, an initial data set is acquired to enable model training. During the acquisition of the initial data set, there must be no significant deterioration in the measurement performance. Ideally, the entire process state space should be covered several times to ensure adequate representativeness of the initial data set. The initial data set should include as many cross-sensitivity parameters (Cp) relevant to the model as possible. iThese include parameters such as the temperature of the measuring tubes, temperature gradients, pressure, density and viscosity of the medium, as well as the substrate temperature, temperature of electronic components of the measuring and operating circuitry, vibration damping and sensor asymmetry, and of course the state parameter Z. The cross-sensitivity parameters Cp i , can be detected using components of the flow meter itself or by means of additional sensors.

[0030] Once sufficient data has been collected for the initial dataset, which can be determined, for example, by observing the passage of a training period (which may last several days or weeks), the initial dataset can be divided into a training set, a validation set, and a test set to train a machine learning algorithm. This algorithm can be of various types, such as Random Forest, Randomized Trees, Ensemble Learning, or advanced deep learning methods like Convolutional Neural Networks (CNNs), Gated Recurrent Units (GRUs), Long Short-Term Memory (LSTMs), or any hybrid combination thereof.

[0031] In addition to model training, an empirical probability distribution function (EPDF) of prediction errors is determined and stored between the state parameter prediction value predicted by the trained model based on the cross-sensitivity parameters and the state parameter measurement value from the test set.

[0032] The model is now ready for use and can leave the training phase to enter the operational phase. In this phase, a new state parameter measurement and associated cross-sensitivity parameter measurements are acquired. This triggers the calculation of a state parameter prediction value and the error between the state parameter prediction value and the state parameter measurement, including determining the probability of this error based on the EPDF learned during the training phase. If, in a comparative test, the probability remains greater than a predefined confidence level (α), it is assumed that the measuring device does not exhibit any deterioration of the state monitored by the state parameter. Therefore, the measurement performance is not affected, and no maintenance, such as recalibration, is required. Otherwise, there is an indication that recalibration is necessary.To limit false alarms, it is recommended to perform a test 300 against a maximum number of acceptable consecutive failed probability conditions before actually triggering a recalibration request 310. In such a case, recalibration would only be triggered if this consecutive number of permissible probabilities less than α is exceeded.

[0033] Furthermore, the dynamics of the condition parameter in the case of monotonic aging over time are autoregressive (not necessarily linear). Therefore, it is advantageous to supplement prior condition monitoring with a time series forecast of either the prediction error or its probability. This would allow the need for recalibration to be anticipated, rather than a change in condition at a specific point in time. The threshold for prediction failure can be either the error corresponding to the confidence level α from the EPDF in the case of error prediction, or the level α itself in the case of error probability prediction. Finally, applying both conditions (probability of prediction error and prediction) enables an efficient method in the case of sudden deterioration or long-term aging. QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. 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

[0000] US 2007 / 028663 A1

[0001] DE 602 18 050 T2

[0001] WO 2012 / 062551 A1

[0001] EP 3 894 802 B1

[0001] DE 10 2010 000 759 A1

[0019]

Images