Methods for detecting the state of a machine using virtual sensors

Abstract

The invention relates to a method for detecting the machine state of a rotating machine, in particular a pump, wherein at least one real sensor is installed at a first measuring position on the machine, which detects at least one first measured quantity at the first measuring position, characterized in that, based on the at least one first measured quantity and a predefined path model, a second measured quantity for a second measuring position of the machine is calculated.

Description

[0001] The invention relates to a method for detecting the machine state of a rotating machine, in particular a pump, wherein at least one real sensor is installed at a first measuring position on the machine, which detects at least one first measured quantity at the first position.

[0002] Sensors are typically used to detect the condition of a machine. The measured values ​​are used, for example, to monitor the machine as well as to control and regulate the machine drive. A classic example is pumps, which incorporate one or more sensors to detect the pump's operating states and use this information for monitoring and control.

[0003] However, the use of sensors results in higher component costs. Furthermore, machine malfunctions are often caused by faulty sensors, which is why efforts are made to operate machines with the fewest possible sensors. Sensors also often cannot be installed at the desired measuring positions because either the available installation space is insufficient, installation is impossible for kinematic reasons, or the prevailing environmental conditions at the measuring point would cause the sensor to malfunction or not function correctly, or its lifespan would be severely compromised.

[0004] Therefore, a solution is being sought to optimally monitor a machine without having to increase the number of sensors used.

[0005] This problem is solved by a method according to the features of claim 1. Advantageous embodiments of the method are the subject of the dependent claims.

[0006] According to the invention, starting from the generic method, it is proposed that, based on the at least one first measured variable of the real sensor and a predefined path model, a second measured variable for a second measuring position on the machine is calculated. The inventive approach thus consists of simulating a virtual sensor at an arbitrary measuring position on the machine. During machine operation, the virtual sensor outputs current measured values ​​for its position-specific measured variable and makes these values ​​available, for example, to a higher-level monitoring, control, or regulation function.

[0007] The calculation of the second measured value for the specific second measurement position of the virtual sensor is ensured using a path model, which is generated, for example, before the machine is commissioned. Using the transfer function, and based on functional equations and with the measured value from at least one real sensor as the input, the measured value at the position of the virtual sensor can be determined. Here, we refer to one real sensor. However, the path model can, of course, be fed with multiple measured values ​​from different real sensors. It is also possible to maintain multiple path models in order to generate more than one virtual sensor and calculate its measured values.

[0008] The method according to the invention offers, for example, the possibility of replacing some of the real sensors required for machine monitoring, control, or regulation with one or more virtual sensors. This reduces the manufacturing costs of the machine and also minimizes its susceptibility to errors due to sensor failures.

[0009] A further advantage of the invention is that, with the aid of the inventive method, virtual sensors can also be simulated at measurement positions that are not accessible to real sensors, or at least not economically accessible. The method theoretically allows locally dependent measured variables to be calculated for any arbitrary measurement position. This improves the overall monitoring of the machine.

[0010] Finally, the method according to the invention also enables improved monitoring for potential malfunctions or wear effects. The pre-generated path model is typically created using a properly functioning and ideally new machine that exhibits no malfunctions and ideally only minimal wear on its machine components. Malfunctions occurring during operation or increasing wear of the machine cause changes in the actual path between the real sensor and the virtual sensor, resulting in deviations from the pre-generated path model. By monitoring the measured value of the virtual sensor, changes in the actual path compared to the path model can be detected, allowing conclusions to be drawn about wear and malfunctions. For example,The calculated second measurement from the virtual sensor is compared with the output of another real sensor, and deviations between the measured value and the calculated value indicate changes to the route.

[0011] As explained above, the plant model is based on a transfer function with the measured value from the real sensor as the input and the measured value from the virtual sensor as the output. Such a transfer function includes, for example, a physical description of a combination of different machine elements and / or their structural-mechanical interaction. Machine elements can include, for example, the machine housing, any installed bearings, the rotating shaft of the machine, any keys, screw connections, or other sealing elements within the machine.

[0012] The first and second measurements can represent identical physical quantities; for example, the real sensor might measure the temperature at a first measurement position, while the virtual sensor calculates the temperature at a second measurement position. However, it is also conceivable that the first and second measurements refer to different physical quantities; for example, the measurement from the real sensor could correspond to a temperature value, while the second measurement from the virtual sensor could correspond to a pressure value or some other physical quantity.

[0013] It is particularly advantageous if the second measuring position is located in an area within the machine where a physical sensor cannot be installed. Such areas are conceivable, for example, within bearings or mechanical seals, especially in the area of ​​their friction surfaces.

[0014] The plant model can be a physical model or a data model. A combination of the aforementioned models to form a hybrid model is also conceivable. The physical model is preferably based on analytical equations that describe the physical relationships between the relevant machine components and their interactions. Instead of an analytical model, a simulation model generated using simulations performed in the test field can also be used. It can also be advantageous if the physical plant model used is a combination of an analytical model and a simulation model.

[0015] The physical relationships, particularly the analytical equations and simulation results, can be captured in the test field by measuring various machine input and output data and correlating their signals. This allows the corresponding measured value to be output as the model's output at the desired position of the virtual sensor. For example, a real sensor could be used at the desired position of the virtual sensor in the test field. Input and output data for parameterizing the path model could include motor currents, acceleration values, mechanical vibration values, acoustic signals, strain measurements, or temperature values.

[0016] When using a data model, a statistical model is preferred, which is generated using statistical data on machines, particularly statistical data from machines of the same or similar type. The statistical data for the data model can be generated, for example, using machine learning methods in the test field or from available data of a fleet of similar or identical machines. In addition to statistical data on the machine or comparable machine types, information on the maintenance history and known machine failures is also considered for parameterizing the data model.

[0017] It is particularly advantageous if the plant model used not only calculates the current measurement value of at least one virtual sensor, but can also output additional information. This additional information includes, for example, forecasts of the remaining service life of the machine or individual machine components. It is also conceivable to output recommendations for optimizing machine operation, in particular specifications for the machine's regulation and control, or maintenance recommendations. Furthermore, it is conceivable to output information about potential damage to the machine or machine components, as well as any wear effects on individual components. This includes, for example, damage such as cavitation, bearing damage, or damage to mechanical seals.

[0018] In addition to the method according to the invention, the present invention also relates to a rotating machine, preferably a pump, and particularly preferably a centrifugal pump, with a control system configured to carry out the method according to the invention. The rotating machine thus offers the same advantages and properties as those already described above with reference to the method according to the invention. Therefore, a repetitive description is omitted.

[0019] In the claimed rotating machine, the control system for executing the process is an integral part of the machine. Alternatively, the invention also includes a system consisting of a rotating machine, preferably a pump, particularly preferably a centrifugal pump, and at least one control system configured to carry out the method of the invention. The control system is not an integral part of the machine, but is arranged separately from the machine and therefore includes a communication interface for exchanging information with the pump.

[0020] Further advantages and features of the invention will be explained in more detail below with reference to an exemplary embodiment and a figure.

[0021] The invention is described below with reference to a specific embodiment in the form of a pump. The single figure shows a centrifugal pump 1 with a pump impeller 2, which is driven via the mechanical shaft 3. The reference numeral 4 designates a real sensor that detects a specific measured value at the location of the axial bearing 7 marked "S".

[0022] For the operation of pump 1, however, it is desirable to obtain additional measured values ​​at different measuring positions that may not be accessible using measurement technology. For example, a measured value should be recorded in the area of ​​the channels of the pump impeller 2. A measurement signal that can be tapped at the pump shaft 3 is also desirable.

[0023] Instead of installing additional physical sensors at the aforementioned measurement positions, virtual sensors 5, 6 are generated according to the invention. Using a precise path model, unique to each virtual sensor 5, 6, which describes the transmission path from the physical sensor 4 to the respective virtual sensor 5, 6, the signal at the virtual sensor 5, 6 can then be calculated based on the signal measured by the physical sensor 4. The transmission path is indicated in the figure by reference numerals 8, 9. Modeling

[0024] To model the path 8, 9, the transfer function from the position of a virtual sensor 5, 6 to a real measuring point 4 is determined. The real sensor 4 can output a measured variable such as local temperature, current, vibrations, or acoustics. The transfer function physically describes a combination of various machine elements (housing 10, bearing 7, shaft 3, keyway, screw connection, seals, etc.) or structural-mechanical interactions of the components that influence the physical quantity to be determined by the virtual sensor 5, 6. Data model versus physical model:

[0025] To utilize virtual sensors 5, 6, a highly accurate path model is required. This can be a physical model, a data model, or a combination thereof. A physical model is typically based on analytical equations or a simulation model such as FEM, CFD, RoM, etc. Unknown parameters are determined through measurements. A data model is based on a statistical approach parameterized with a large number of input and output combinations. The modeling can be divided into the following three approaches: 1. Physical model 2. Data model (test bench) 3. Data model (cloud - fleet data)

[0026] Approach 1 is based on an analytical model, which may be parameterized in the test field. Approach 2 does not use an analytical model, but rather a black box that is parameterized using extensive test signals in the test field. Approach 3 uses data from the fleet, which is provided via a cloud interface. The fleet comprises, for example, as large a number as possible of identical or at least similar pumps.

[0027] The advantage of approach 2 is that it allows for a very precise analysis of the cases under investigation. The advantage of approach 3 is that the very large number of machines allows for a significantly higher degree of variation to be modeled. Furthermore, approach 3 captures real-world customer use cases rather than laboratory conditions.

[0028] By combining all three approaches, the amount of data required to create the data model can be significantly reduced. In this case, all existing physical relationships do not need to be learned from scratch using data-driven methods; instead, these relationships are derived from physical models. With a so-called hybrid model, consisting of a physical model and a data model, operational anomalies and errors can also be understood much more effectively. Parameterization of the route model:

[0029] The system model can be analytical or statistical. A simulation model (FEM, CFD, RoM, FMU, etc.) can also be used. The system model requires measured input and output data for parameterization. Input data includes time-based sensor data such as motor current, accelerations, vibrations, acoustic signals, strains measured with strain gauges, and temperature. Statistical data (metadata) from the pump unit can also be used. Maintenance history and identified failures are also important for parameterizing the model.

[0030] This parameterized model outputs current signals at the positions of the virtual sensors 5 and 6. Furthermore, it can generate remaining service life predictions or recommendations for action. These recommendations can be used to optimize pump operation. Additionally, damage such as cavitation, bearing damage, or damage to the mechanical seals can be detected.

[0031] The following practical applications of the method according to the invention are conceivable and advantageous: Case 1: Replacement for real sensors

[0032] Every sensor is subject to the potential risk of failure and represents a cost. Replacing a physical sensor with a virtual one helps to avoid these disadvantages. Case 2: Additional sensors:

[0033] It is not possible to place an unlimited number of sensors. If a well-parameterized route model is available, virtual sensors can be applied at any point, even at locations that were not previously examined with a real sensor (during parameterization). Case 3: Machine diagnostics

[0034] By using two real sensors and one virtual sensor, the plant model can be verified during operation. The condition of the machine elements influences the actual transfer function between two measuring points, so that deviations between the values ​​of the virtual sensor and the real sensor can indicate fault conditions such as wear, cavitation, etc.

[0035] Fault detection using the pipeline model can be performed either at the edge or in the cloud. Edge-based means that the associated algorithms are executed on a purchased component, such as a pump unit, sensor, or gateway. Cloud-based means that the component sends data to the cloud, and the algorithm itself runs in the cloud.

[0036] In a specific implementation, a vibration sensor, for example, could be used as the real sensor 4, mounted on the bearing support (at a defined location) of the bearing 7. To generate the path model, the pump 1 is measured in the laboratory, whereby measurement data, in particular the vibration signals of the real sensor 4, are recorded at various load points of the pump and / or under artificially generated balancing during operation of the pump.

[0037] The vibration signals are then analyzed. This can include classification into healthy and faulty hydraulic states (partial load, overload, cavitation). Classification into healthy and faulty mechanical states (balance quality, misalignment) is also performed. The analysis also takes into account possible customer-specific boundary conditions, such as different installation situations. The classified data, along with supplementary information, then forms the input for the data-driven model of the transmission line.

[0038] The next step involves modeling the transmission path, generating mechanical and hydraulic models based on the classification. Mechanical models simulate, among other things, stiffness, damping, installation, preload, bolt tightening torques, pipe connections, and pump mounting orientation. Hydraulic models simulate hydraulic states such as partial load, rated load, overload, and cavitation. The model is also capable of representing interactions (e.g., mechanical excitation by hydraulic forces).

[0039] The resulting plant model is capable of outputting measured values ​​for the simulation of virtual sensors during normal machine operation, based on the vibrations occurring and measured by the vibration sensor at one or more measurement positions. Furthermore, the model is capable of detecting four fault conditions during operation based on the measured vibrations of sensor 4.

Images