Method for operating a centrifuge, computer-implemented method for training a machine learning tool, computer program and centrifuge

Abstract

Method for operating a centrifuge that has at least one rotatable bowl (1) with a pressure controller (91), configured to set a level of a back-pressure inside the bowl (1), and that has an inlet pipe (40) and that has at least one liquid outlet (13), with which, during operation, a flowable starting product (P) is separated into different liquid and / or solids phases in the centrifugal field of the rotating bowl (1), the method comprising at least the following computer-implemented steps for recognising an operating state (511): 100 Sampling signal data (101) with at least one sensor (22) arranged on the centrifuge, the signal data (101) comprising a discrete-time structure-borne sound signal (111) representing a temporal course of a structure-borne sound signal detected using at least one structure-borne sound sensor (22); 200 Transforming the discrete-time structure-borne sound signal (111) to the frequency domain, to obtain a set of frequency spectrum data (211); 300 Inputting the set of frequency spectrum data (211) into a computerimplemented machine learning tool (301) executed on a control device (9); 400 Evaluating the set of frequency spectrum data (211) through a computer-implemented algorithm (3011) of the machine learning tool (301), hence determining a cavitation score (411), and providing the cavitation score (411) to the control device (9) of the centrifuge, by the computer-implemented algorithm (3011) of the machine leaning tool (301); 500 Determining of an operating state (511) of the centrifuge based on the cavitation score (411) by the control device (9); 600 Regulating the back-pressure inside the bowl (1) of the centrifuge through the control device (9) by sending a control signal (611) that is dependent on the operating state (511) from the control device (9) to the pressure controller (91).

Description

[0001] The present invention relates to a method for operating a centrifuge with cavitation monitoring, a computer-implemented method for training a machine learning tool, a computer program and a centrifuge.

[0002] Centrifuges can be used to separate a starting product into several phases, a light and a heavy phase, for example. Cavitation of a liquid inside a centrifuge can be a source of substantial damage to the centrifuge parts and should thus be monitored or prevented or mitigated with counter-measures.

[0003] EP 3 485 979 B1 discloses a method for detecting the operating state of a centrifuge by detecting a structure-borne sound signal using a structure-borne sound sensor near to the bowl of the centrifuge. There, a temporal course of the signal is detected and converted into the frequency domain by means of a Fast Fourier Transform. The resulting frequency spectrum is compared to one or more known spectra to find deviations that might indicate an operating state of the centrifuge. According to EP 3 485 979 B1 this method can be used to monitor the centrifuge and detect cavitation or a threat of cavitation occurring.

[0004] To mitigate cavitation or prevent it from happening inside centrifuges, a back-pressure inside the centrifuge bowl, typically set on an output line of the centrifuge, is commonly increased to a level suitably high to stop cavitation bubbles from forming. EP 3 485 979 B1 discloses increasing the back-pressure in response to a cavitation at a centripetal pump of the centrifuge.

[0005] A problem with the known cavitation monitoring means for centrifuges is their limited reliability. This necessitates the use of back-pressure levels that are substantially higher than is strictly necessary for suitable operation, resulting in undesirably high energy consumption of the centrifuge and a resulting efficiency reduction, while the monitoring system also increases the manufacturing costs of the machines. Another problem is that cavitation analysis of centrifuges is typically carried out at frequencies that are well into the medium frequency (MF) band, which requires sophisticated spectrum analysers, driving up setup cost substantially.

[0006] It is thus desirable to develop cheap, reliable means for the monitoring of centrifuges and to prevent cavitation with optimised energy consumption.

[0007] The present invention solves this problem with a method according to claim 1, a computer-implemented method for training a machine learning tool according to claim 18, a computer program according to claim 20 and a centrifuge according to claim 21.

[0008] According to the present invention, the method for operating a centrifuge that has at least one rotatable bowl with a pressure controller, configured to set a level of a back-pressure inside the bowl, and an inlet pipe and that has at least one liquid outlet and / or one solids outlet, with which, during operation, a flowable starting product is separated into different liquid and / or solids phases in the centrifugal field of the rotating bowl, comprises at least the following computer-implemented steps for recognising an operating state: 100 Sampling signal data with at least one sensor arranged on the centrifuge, the signal data comprising a discrete-time structure-borne sound signal representing a temporal course of a structure-borne sound signal detected using at least one structure-borne sound sensor; 200 Transforming the discrete-time structure-borne sound signal to the frequency domain, to obtain a set of frequency spectrum data; 300 Inputting the set of frequency spectrum data into a computer-implemented machine learning tool executed on a control device; 400 Evaluating the set of frequency spectrum data through a computer-implemented algorithm of the machine learning tool, hence determining a cavitation score, and providing the cavitation score to the control device of the centrifuge, by the computer-implemented algorithm of the machine leaning tool; 500 Determining of an operating state of the centrifuge based on the cavitation score by the control device; 600 Regulating the back-pressure inside the bowl of the centrifuge through the control device by sending a control signal that is dependent on the operating state from the control device to the pressure controller.

[0009] The steps 300 and 400 enable a particularly accurate and reliable, cost-effective analysis of the sampled signal data and enable particularly reliable determination of the operating state of the centrifuge based on the cavitation score provided by the computer-implemented machine learning tool. This increases the reliability of the cavitation monitoring of the centrifuge such that the back-pressure inside the bowl can be regulated more appropriately. In particular, the present invention enables a minimisation of the back-pressure to a level that only just avoids cavitation and hence eliminates the need for a significant safety margin in the back-pressure. The energy consumption of the centrifuge is thus optimised and its operating efficiency increased.

[0010] By making the cavitation detection mechanism more reliable, it can also be possible to work with lower signal-to-noise ratios when detecting the structure-borne sound signals, thus improving cavitation detection at lower frequencies, resulting in an additional cost reduction of the system by requiring simpler sensor read-out technology. The method steps 100 to 600 are typically executed periodically / continuously, i.e. for continuous operation and cavitation monitoring of the centrifuge.

[0011] Preferably, the machine learning tool, and hence the computer-implemented algorithm, was trained using supervised learning to yield especially reliable results, but reinforcement or unsupervised learning techniques could also be employed.

[0012] The machine learning tool can have been trained using training data that comprises a multitude of sets of training frequency spectrum data, acquired by transforming a multitude of temporal courses of training structure-borne sound signals acquired on the centrifuge, or another centrifuge of essentially identical build, into the frequency domain, as an input and respective training cavitation scores as an output. A greater amount of training data can increase the accuracy of predictions by the machine learning tool's algorithm, in this case yielding an overall improvement of the control of the centrifuge and thus its energy efficiency. The structure-borne sound data that is processed, i.e. the structure borne sound signal that is transformed, is the detected, meaning sampled, discrete-time structure-borne sound signal in each of the variants described in this application, including the validation and test data, as well as the data mentioned with respect to the figures. The back-pressure setting and the speed of rotation of the bowl can also be included in the training data in an embodiment.

[0013] Preferably, the structure-borne sound signals are acquired over a frequency range from 0 kHz (Kilohertz) to 30 kHz and most preferred from 10 kHz to 20 kHz for particularly reliable detection. Additionally or alternatively, the structure-borne sound signals are acquired over a frequency range up to 2 kHz or up to 10 kHz. This allows optimisation of the frequency range for the machine learning tool, for example based on an evaluation for a number of training ranges done using test data. Optionally it can also be acquired up to 1 or more MHz (Megahertz), although for cost reasons lower frequencies are preferred.

[0014] The training cavitation scores can essentially be seen as label cavitation scores as an answer key for the training data. The data can be labelled manually or automatically.

[0015] According to one embodiment, the labels are assigned manually to structure-borne sound signal data collected on the centrifuge, or another centrifuge of essentially equivalent build, before the training of the machine learning tool. In another preferred embodiment, the training cavitation scores are set based on data collected using a cavitation detection device that is only connected to the centrifuge for the purpose of the collection of the training data. This device can be a highly precise vibration sensor, for example one operating into the MHz range. Typically this also samples temporal courses of the structure-borne sound signal for different centrifuge settings, although up to the higher frequency, which is then likewise transformed into the frequency domain, from which thus obtained frequency spectra clear assessments of the cavitation risk are possible according to the art and hence cavitation scores can be assigned to these high bandwidth spectra simply. These scores are then correlated to the spectra collected with the structure-borne sound sensor in the lower frequency range.

[0016] According to an especially preferred embodiment, the training structure-borne sound signals are acquired by the at least one structure-borne sound sensor and the training cavitation scores (the labels) are ones detected simultaneously using the cavitation detection device that detects structure-borne sound signals up to a higher frequency than the structure-borne sound sensor. This can be visualised as detecting the training structure-borne sound signals in the same manner as in step 100 and / or transforming them in the same manner as in step 200, and then assigning a training cavitation score that was established using the cavitation detection device for the same temporal course to each set of frequency spectrum data. Figuratively, a vector of frequency spectrum data collected using the structure-borne sound sensor and one of cavitation scores detected using the cavitation detection device are concatenated to create this set of training data. This results in particularly accurate training data, hence improving the machine learning tool once trained. In particular, the machine learning tool of the present invention can thus be trained particularly effectively using a more precise, but also more expensive / complex cavitation detection device (for example the MHz-range vibration sensor), while its normal operation can carry out the cavitation detection using a cheaper and simpler sensor that has enhanced insight into cavitation effects on its read-out signal as a result of the machine learning tool. The same is also possible for the validation and / or test data.

[0017] Particularly, the machine learning tool can have been trained further, or tuned, using validation data that comprises a multitude of sets of validation frequency spectrum data, acquired by transforming a multitude of temporal courses of validation structure-borne sound signals acquired on the centrifuge, or another centrifuge of essentially identical build, into the frequency domain, as an input and respective validation cavitation scores as an output.

[0018] Preferably, the training and / or validation structure-borne sound signals are acquired over a frequency range of up to 10 kHz, up to 20 kHz, from 10 kHz to 20 kHz, or up to 1 or more MHz.

[0019] It is particularly desirable that the machine learning tool was verified, or evaluated, using test data that comprises a multitude of sets of test frequency spectrum data, acquired by transforming a multitude of temporal courses of test structure-borne sound signals acquired on the centrifuge, or another centrifuge of essentially identical build, into the frequency domain, as an input and respective test cavitation scores as an output.

[0020] Preferably, the test structure-borne sound signals are acquired over a frequency range of up to 10 kHz, up to 20 kHz, from 10 kHz to 20 kHz, or up to 1 or more MHz.

[0021] Any of the temporal courses of structure-borne sound signals acquired on the centrifuge that are listed above are preferably transformed into the frequency domain by means of a Fast Fourier Transform (FFT), to obtain the respective frequency spectrum data, i.e. a multitude of frequency spectra. A temporal course to be transformed can be a window of time, where a series is recorded and one window follows the next (i.e. each window comprising samples n to n+[samples per window-1], with n being the first sample identifier of each window), or it can be a series of windows where the first sample is replaced upon recording of another, in a first-in-first-out (FIFO) mechanism.

[0022] In the context of the present application it is clear that the sensor and / or the control device comprise means for sensor-readout. The structure-borne sound sensor can directly comprise means to provide a digitised signal to the control device, or it can provide an analogue signal that is digitised by a converter comprised in the control device (which includes an implementation using a discrete converter (typically an analogue-to-digital converter (ADC))). Frequency-based sensor read-out can also be used, but is more complex.

[0023] Acquiring either or all of the training, label, validation and the test structure-borne sound data on the centrifuge of the present inventive method (i.e. the centrifuge actually used for the separation operation) or on the centrifuge according to the present invention has the advantageous effect that the machine learning tool will be particularly accurate and will be able to better reflect manufacturing variations. However, alternatively acquiring either or all of the training, label, validation and the test structure-borne sound data on a, meaning another, centrifuge of essentially identical, i.e. equivalent, build has the advantage that the machine learning tool may be trained once and subsequently computer-implemented on the control devices of a multitude of centrifuges that is stored on, reducing the cost of training and hence the centrifuge.

[0024] According to one variant, the machine learning model was trained on a centrifuge of essentially equivalent build to the centrifuge of the method of the present invention and then retrained on the centrifuge of the method of the present invention. This enables the implementation of a particularly accurate machine learning tool at decreased computational cost, as retraining the algorithm to a desired accuracy can be done faster and with smaller training datasets than training an algorithm, i.e. also the machine learning tool, from scratch for every centrifuge.

[0025] The training of the machine learning tool using supervised learning is conducted using typical methods according to the state of the art. In particular, the training data is input into the, at his point before the tool is trained, yet untrained machine learning tool, where the untrained machine learning tool receives the training frequency spectrum data, and evaluates it using an iterative optimisation algorithm or a generative training algorithm, yielding output cavitation scores which the machine learning tool then compares to the training cavitation scores that are defined as part of the training data. According to this result, the machine learning tool adjusts its function if it identifies that the algorithm does not yield the output cavitation score defined by the training data for a particular input, to improve the fit. This process is then repeated iteratively until a preset maximum error margin is achieved when a set of training frequency spectrum data is evaluated by the computer-implemented algorithm to yield a cavitation score, meaning that this cavitation score should be within a certain maximum range of the cavitation score assigned to the dataset, i.e. the label of the set of frequency spectrum data.

[0026] This process is then preferably repeated for the validation data, i.e. the sets of validation frequency spectrum data is input into the machine learning-tool, whose algorithm has at this point been trained using the training data, and the algorithm's parameters are then iteratively adjusted until the algorithm also yields the correct output validation cavitation score for a respective input of a set of validation frequency spectrum data within a preferred error margin.

[0027] In an especially preferred embodiment, this, trained, algorithm can be evaluated for a set of test data, comprising multiple sets of test frequency spectrum data with respective output test cavitation scores. When a performance evaluation of the algorithm yields insufficiently low discrepancies between cavitation scores output by the algorithm for given input test frequency spectrum data and the respective label of the data, i.e. the test cavitation score, the model has to be trained further using more training data that was either preserved at the start of the training cycle or has to be gathered subsequently. It is also possible to set up the machine learning tool for continuous improvement of the algorithm in the computer-implementation in-situ, where for example a feedback can be provided to the machine learning tool when a cavitation occurs in the system that it has failed to predict or prevent from happening by controlling the pressure correctly through outputting the appropriate cavitation score.

[0028] In a particularly preferred embodiment, a cavitation score within a first pre-set value range corresponds to an operating state "imminent cavitation" that indicates that an occurrence of a cavitation inside the bowl is imminent, and wherein a cavitation score within a second pre-set value range corresponds to an operating state "low cavitation risk" that indicates that an occurrence of a cavitation inside the bowl is unlikely.

[0029] It is possible that the value range of the cavitation score ranges from 0 to 1 (i.e. 0 % to 100 %), with a value of 0 corresponding to "no cavitation", with a value from 0.1 to 0.3 corresponding to the operating state "low cavitation risk" and a value from 0.7 to 0.9 indicating "cavitation imminent". Here a score of 1 could indicate an operating state "cavitation", the occurrence of which would also be countered by increasing the back-pressure, but could also be reacted upon by terminating the operation of the centrifuge. These ranges only serve as an example. It is also thinkable to use a "single" threshold value for the cavitation score, wherein the operating states can respectively be defined on either side of the score, for example such that below the threshold the back-pressure is reduced and above the threshold it is increased. A higher value could likewise also correspond to a low cavitation risk, while a low value could also represent a high risk. The important aspect here is that the control device is configured such that the back-pressure is reduced when the cavitation score indicates that no cavitation is likely to occur and that it is increased when the computer-implemented machine learning tool identifies that cavitation is likely to occur.

[0030] The setting of the value range can be designed with regards to a desired precision of the cavitation monitoring and the corresponding back pressure control response. For example, a larger range of values might be used, if very fine control of the pressure controller is desired. This way, an embodiment could be implemented, where each cavitation score corresponds to an operating state that is assigned to a pre-set back-pressure level, such that the control device is configured to control the pressure controller such that it will set this back-pressure upon a respective cavitation score being observed.

[0031] The control device can be configured to increase the back-pressure inside the bowl of the centrifuge when the operating state "cavitation imminent" is detected, by sending a control signal to increase the back pressure inside the bowl to the pressure controller, and wherein the control device is configured to decrease the back-pressure inside the bowl of the centrifuge when the operating state "no cavitation" (or possibly also "low cavitation risk") is detected, by sending a control signal to decrease the back pressure inside the bowl to the pressure controller. When an operating state "high cavitation risk" is detected, the back pressure can be either maintained or increased, depending on the desired certainty of cavitation avoidance.

[0032] The pressure controller can be configured to increase or decrease the back-pressure inside the bowl depending on the control signal from the control device, typically on the outlet line of the centrifuge. The pressure controller may, for example, comprise a solenoid valve and / or a pump or a compressed air supply, which is regulated by a control unit that can be part of the pressure controller or the control device, wherein the control signal is used to control the operating state of the solenoid valve and / or the pump or the compressed air supply. The pressure control is preferably implemented by a discharge valve on each phase of the centrifuge. The Pressure increases from the outlet to the centripetal pump through throttling up to an overflow of a heavy phase.

[0033] It is preferred that each of the structure-borne signals is acquired using the at least one structure-borne sound sensor, or at least one essentially identical structure-borne sound sensor on another centrifuge of essentially identical build, when the latter embodiment is implemented, where the machine learning tool is trained on another device. An embodiment encompassing multiple structure-borne sound sensors at different locations on one centrifuge is also possible, where the machine learning tool either evaluates and was trained for each of the signals, or where the signals are evaluated, e.g. averaged or convoluted and / or filtered and transformed (or first transformed and then multiplied), to yield a single, more accurate dataset representing a sort-of "average" frequency response of the whole centrifuge that is then applied to the machine learning tool for training and later for implementation of steps 100 to 400 of the method according to the invention.

[0034] Step 400 can include the evaluation of other components of the signal data. For example, the signal data, here the data, evaluated by the machine learning tool in step 400 can include a course of back-pressure settings for the temporal course of the structure-borne sound signal for a particularly accurate determination of the cavitation score. In the same way, back-pressure data can be included in the training data the machine learning tool was trained with.

[0035] The data for teaching the machine learning model was preferably collected for one or more centrifuge setups and the machine learning model was thus preferably trained and set up for one or more centrifuge setups, for example for different rotation speeds of the bowl. The training data could, for example be data that was collected using a known electrolyte inside the centrifuge at a range of centrifuge settings using a structure-borne sound sensor. Preferably this structure-borne sound sensor is one installed permanently on the centrifuge, but another sensor may also be used for gathering the training data, before installing the sensor intended for in-situ use of the centrifuge.

[0036] In an embodiment, the signal data can further comprise (ambient and / or internal) temperature data, humidity data, process data like bowl speed or processed fluid, or pressure data, which the machine learning tool was preferably also trained with, in line with the training outlined above, where the training data also comprises equivalent training, label, validation and the test signal data of the cited types that was acquired in the same way. Alternatively, or additionally, this data is used to increase the reliability of the machine learning tool and hence the inventive method. The additional data can particularly be used to compensate for variations in either of the factors monitored thus, especially temperature or possibly ambient air pressure.

[0037] The additional monitoring could also be employed to configure the control device such that it will either output a warning, or increase the back-pressure above a certain safety margin, and in this case hence proceed by switching the operation of the centrifuge to an operation according to the state of the art, when any of the additional monitored conditions, e.g. temperature or ambient air pressure, vary such that a reliable use of the machine learning tool, and hence the regulation of the back-pressure according to step 600, is no longer, or temporarily, not possible. This can result in a particularly cavitation-proof design or operation of the centrifuge.

[0038] As is typical for the training of machine learning tools, the training, label, validation and the test structure-borne sound data or other signal data used for these purposes should preferably have the same, or almost the same probability distribution for optimal training and validation of the machine learning tool.

[0039] The control device can be configured to issue a warning message or alert when the operating state "cavitation imminent" is detected. This can be sent to the pressure controller or a device, that can be part of the pressure controller, which sends the alert to an operator of the centrifuge and can be accompanied by the suspension of its operation.

[0040] In a preferred embodiment, the computer-implemented machine learning tool has an algorithm based on a classification algorithm. The algorithm can in particular be based upon an artificial neural network, nearest neighbours, linear regression, naïve Bayes, a decision tree, for example random forest, or a support vector machine. Using an autoencoder is also possible; perceptron could be used for binary pressure control based on only two operating states. Preferably the algorithm simply employs only one of these topologies, using the structure-borne sound signal (specifically its transformed version) as the inputs, and having been trained using training, validation and / or test data as described above. While the machine learning algorithm needs to be computer-implemented as required by the present invention, the choice of a particular algorithm is not absolutely essential, but should rather be made based on a desired reliability of the cavitation monitoring and the training data available. To this end, the machine learning tool can also be implemented by first training multiple algorithms with the training data and / or the validation data, after which each algorithm is evaluated using the test data, with the algorithm yielding the lowest error being chosen for ultimate computer-based implementation on the control device for the execution of the inventive method.

[0041] In another preferred embodiment, the machine learning tool was trained by evaluating the signal data, in particular the training, validation and test data as referenced above on a selection of the listed algorithm types and the most effective algorithm was selected.

[0042] According to the present invention, a computer-implemented method for training a machine learning tool is also provided. The method provides a cavitation score based on a structure-borne sound signal detected on a centrifuge, the method including a step of training the machine learning tool in an untrained state using training data that comprises a multitude of sets of training frequency spectrum data, acquired by transforming a multitude of temporal courses of training structure-borne sound signals acquired on the centrifuge, or another centrifuge of essentially identical build, into the frequency domain, as an input and respective training cavitation scores as an output.

[0043] Preferably, the method comprises a step of tuning the machine learning tool using validation data that comprises a multitude of sets of validation frequency spectrum data, acquired by transforming a multitude of temporal courses of validation structure-borne sound signals acquired on the centrifuge, or another centrifuge of essentially identical build, into the frequency domain, as an input and respective validation cavitation scores as an output.

[0044] A computer program including a set of instructions that, when executed on a control device, cause the control device to implement the method according to any one of the embodiments of the present invention is also comprised in the present invention. The control device can in particular be a processor, specifically it can be a computer, an industrial computer (IPC), a programmable logic controller (PLC) or a microcontroller or it can comprise an artificial intelligence (Al) accelerator.

[0045] The present invention further includes a centrifuge, to which a structure-borne sound sensor is attached and which has at least one rotatable bowl with an inlet pipe and at least one liquid outlet and / or one solids outlet, with which a flowable starting product can be separated into different liquid and / or solid phases in the centrifugal field of the rotating bowl during operation, wherein the centrifuge further comprises a control device which can detect a structure-borne sound signal of the centrifuge by means of the structure-borne sound sensor and which comprises a processing device which implements the method for operating the centrifuge according to the present invention.

[0046] In a particularly preferred embodiment, the control device is configured to control the back pressure, wherein a cavitation score within a first pre-set value range corresponds to an operating state "imminent cavitation" that indicates that an occurrence of a cavitation inside the bowl is imminent, and wherein a cavitation score within a second pre-set value range corresponds to an operating state "low cavitation risk" that indicates that an occurrence of a cavitation inside the bowl is unlikely.

[0047] According to an even more preferred embodiment, the control device is also configured to increase the back-pressure inside the bowl of the centrifuge when the operating state "cavitation imminent" is detected, by sending a control signal to increase the back pressure inside the bowl to the pressure controller, and wherein the control device is configured to decrease the back-pressure inside the bowl of the centrifuge when the operating state "cavitation unlikely" is detected, by sending a control signal to decrease the back pressure inside the bowl to the pressure controller.

[0048] The processing device can include one or more of an industrial PC, a programmable logic controller, a microprocessor, a microcontroller, a field-programmable gate array (FPGA) or an application-specific integrated circuit (ASIC). To increase the execution speed of the computer-implemented method, the processing device can additionally comprise an artificial intelligence accelerator that accelerates the implementation of the machine learning algorithm, in particular when implemented together with or as part of a microprocessor or an ASIC.

[0049] It is also possible to implement the machine learning tool on a remote processing device that is in communication with a processor, i.e. a PLC or similar, on the centrifuge. A wired or wireless communication are both conceivable for this implementation. The machine learning tool could thus be cloud-based and, for example, be accessible by multiple centrifuges at once. The sampled data can here be sent to a remote processing unit, where the transformation and evaluation of the data happens, and the cavitation score returned to the control device. This would also enable easy updating of the machine learning tool to quickly take advantage of improvements to the algorithm made through further, successive training over time. The centrifuge could also be "self-learning", such that a detection of occurring cavitation that the algorithm has not successfully foreseen, could be used to avoid a similar reoccurrence. This way the method could be improved especially in unstable environments and the centrifuge or the method hence be more universally deployable.

[0050] In a preferred embodiment, the structure-borne sound sensor can be a vibration sensor, preferably an accelerometer. Especially in frequency ranges under 20 kHz, vibration sensors that can be used as structure-borne sound sensors are readily available at low cost, with simple and cost-efficient read-out and drive electronics.

[0051] The operation of the centrifuge for separation is otherwise implemented as known in the art. For additional information on this, EP 3 485 979 B1 is incorporated by reference, in particular paragraphs

[0022] to

[0030] that outline its functionality.

[0052] Preferably the centrifuge can be a separator. The invention can alternatively also be used on nozzle separators or on separators without an outlet for solids.

[0053] Below, aspects of the present invention are described in further detail with reference to the figures. It is shown in: Fig.1a schematic section view of a centrifuge according to the invention; Fig. 2a, ba flow chart of the inventive method for operating a centrifuge; Fig. 3a flow chart of the functionality of the machine tool; Fig. 4a flow chart of a method of training of the machine tool;

[0054] Fig. 1 shows a centrifuge according to the present invention - here designed as a separator - for clarifying solid-containing, flowable starting products P from solid matter with a rotatable bowl 1 with a preferably vertical axis of rotation D.

[0055] The processing of the starting product P takes place in continuous operation. The separator here is a self-emptying separator.

[0056] This means that the feed of the starting product P - which is a flowable suspension - is continuous and the discharge of at least one clarified liquid phase, called clear phase L, is also continuous. The bowl 1 of the centrifuge, in the preferred implementation as a self-emptying separator, has a discontinuous solids outlet, whereby the solids separated from the starting product P by clarification are discharged at intervals by opening and re-closing outlet nozzles or solids discharge openings 5 are removed. Each of the phases resulting from this separation can - but does not necessarily have to - form a valuable material phase to be recovered.

[0057] The invention can alternatively also be used on nozzle separators or on separators without a solids outlet. It can also be used on separators that do not operate continuously in batch mode.

[0058] The bowl 1 has a bowl bottom 10 and a bowl top 11. It is further preferably surrounded by a hood 12. The bowl 1 is also mounted on a drive spindle 2, which is rotatably mounted and can be driven by a drive motor. The bowl 1 itself is rotatable and forms an essential part of the rotating system of the centrifuge, but it also has individual elements protruding into it which do not rotate during operation. Thus, the bowl 1 has a product inlet 4 through which the starting product P is fed into the bowl 1. The product inlet 4 opens into an inlet pipe 40, which here is designed as a pipe that does not rotate with the rotating system - i.e. does not rotate during operation - which projects into the bowl from above and is aligned coaxially with the axis of rotation D. Alternatively, it would also be conceivable for the inlet pipe 40 to protrude into the bowl from below (with a correspondingly different structural design - not shown here).

[0059] The bowl 1 further comprises at least one outlet 13 - which is designed here as a peeling disk or as a centripetal pump - which serves to drain a clear phase L from the bowl 1. The outlet 13 can also be implemented constructively in a different way or by other means. It is also conceivable, as an alternative or in addition to the clarification of solids, to separate the starting product P into two liquid phases of different densities. This requires an additional fluid drain - for example, another centripetal pump. The centripetal pump thus also forms a component of the centrifuge which does not rotate with the actual bowl 1 during operation but remains stationary.

[0060] The bowl 1 preferably has a disc stack 14 made up of axially spaced separating discs. A solids holding space 8 is formed between the outer circumference of the disc stack 14 and the inner circumference of the bowl 1 in the region of its largest inner diameter. Solids (solid matter) which are separated from the clear phase in the area of the disc stack 14 are collect in the solids holding space 8, from which the solids can be discharged from the bowl 1 via solids discharge openings 5. For this purpose, the solids discharge openings 5 can be opened and closed by means of a sliding piston 6 (or valve), which is arranged in the bowl bottom 10 and can be moved therein parallel to the axis of rotation (in particular vertically). When the solids discharge openings 5 are open, the solid matter are discharged from the bowl 1 into a solids catcher 7. The solids holding space 8 in the bowl 1 has a defined solids volume.

[0061] The bowl 1 has an actuating mechanism for moving the sliding piston 6. Here, it comprises at least one supply line 15 for a control hydraulic fluid such as water and a valve arrangement 16 in the bowl 1 and further elements outside the bowl 1. Thus, the supply of the control fluid such as water is made possible via a solenoid valve block 17 arranged outside the bowl 1, which is assigned to a supply line 19 for the control fluid, here preferably water, arranged outside the bowl 1, so that for a solids emptying of the solids by releasing the valve arrangement 16, the control fluid can be introduced into the bowl 1 or, conversely, the supply of control fluid can be interrupted in order to move the sliding piston 6 accordingly in order to release the solids discharge openings 5.

[0062] At least one structure-borne sound sensor 22 is arranged on the bowl 1 (here, for example, on a ring of the bowl 1, where the outlet 13 and inlet pipe 40 are connected) and is configured to sample structure-borne sound data. This structure-borne sound sensor 22 is designed as a sensor device for measuring a temporal course of a structure-borne sound signal. The structure-borne sound sensor 22 is preferably an accelerometer, particularly one that can sample a structure-borne sound signal of the system at frequencies from 0 kHz to 20 kHz.

[0063] A back-pressure is controlled by a pressure controller 91, here on the outlet line 13. The pressure controller 91 is in communication with the control device 9 via a communication link 92 and preferably comprises a valve, e.g. a solenoid valve, to set the back-pressure.

[0064] The control device 9 executes a method according to the invention. This method is illustrated in Fig. 2a and 2b.

[0065] Figure 2a shows a schematic diagram of a method for operating a centrifuge (either can be according to the present invention in this case), here the centrifuge of Fig. 1. The centrifuge has the at least one rotatable bowl 1 with the pressure controller 91, configured to set a level of a back-pressure inside the bowl 1, and the inlet pipe 40 and having the at least one liquid outlet and / or one solids outlet 13, with which, during operation, the flowable starting product P is separated into different liquid and / or solids phases in the centrifugal field of the rotating bowl 1, the method comprising at least the following computer-implemented steps, shown also in Fig. 2b and Fig. 3 for recognising an operating state: 100Sampling signal data 101 with at least one sensor 3, 20, 22 arranged on the centrifuge, the signal data 101 comprising a discrete-time structure-borne sound signal 111 (in Fig. 2b and 3) representing the temporal course of the structure-borne sound signal detected using at least one structure-borne sound sensor 22 that is provided on the bowl 1 of the centrifuge here;200Transforming the discrete-time structure-borne sound signal 111 to the frequency domain, to obtain a set of frequency spectrum data 211 (in Fig. 2b and 3), preferably by the control device 9 (preferably comprising a data processing and a data storage unit) of the centrifuge or by the sensor 22 or a separate sensor control in communication with the control device 9 and the structure-borne sound sensor 22;300Inputting the set of frequency spectrum data into a computer-implemented machine learning tool 301 (in Fig. 3), which is here executed, i.e. implemented on the control device 9, for example as a computer program, i.e. a software, that is pre-loaded onto the control device 9 or as hardware comprised in the control device 9;400Evaluating of the set of frequency spectrum data through a computer-implemented algorithm 3011 (in Fig. 3) of the machine learning tool 301, hence determining a cavitation score 411 (in Fig. 2b and 3), and providing the cavitation score 411 to the control device 9 of the centrifuge, by the computer-implemented algorithm 3011 of the machine leaning tool 301;500Determining of an operating state 511 (a to d in Fig. 2b and 3) of the centrifuge based on the cavitation score 411 by the control device 9;600Regulating the back-pressure inside the bowl 1 of the centrifuge through the control device 9 by sending a control signal 611 (a to d in Fig. 2b and 3) that is dependent on the operating state from the control device 9 to the pressure controller 91.

[0066] This method is intended for continuous operation of the centrifuge, i.e. that sampling of structure-borne sound signals, their transformation into the frequency domain (typically by FFT) and the evaluation steps 400 to 600 are carried out periodically while the centrifuge is running.

[0067] Fig. 2b shows the operation of the steps 500 and 600 in particular detail. As part of step 500, the control device determines an operating state depending on a cavitation score 411 established through the computer-implemented algorithm in step 400. This algorithm can be a trained machine learning algorithm according to the state of the art, in particular the machine learning tool employs a classification algorithm as described above, to assign specific frequency spectra (or just frequency responses observed in these, although whole spectra are preferred for training the algorithm) to respective cavitation scores. Each frequency spectrum in step 200 preferably represents consecutive windows of time or windows of time where the discrete datapoints are replaced over time according to a first-in-first-out (FIFO) mechanism.

[0068] As Fig. 2b illustrates, the steps are carried out consecutively. In Step 500 an operating state of the centrifuge is determined by the control device 9 depending on the cavitation score 411 provided by the algorithm, i.e. output from the machine learning tool 301. An operating state "low cavitation risk" 511a is determined when the cavitation score 411 is within a pre-set range, for example 0.1 to 0.3, an operating state "high cavitation risk" 511b is determined when the cavitation score 411 is within a second pre-set range, for example 0.4 to 0.6, an operating state "cavitation imminent" 511b is determined when the cavitation score 411 is within a third pre-set range, for example 0.7 to 0.9, and an operating state "cavitation" 511b is determined when the cavitation score 411 is within a fourth pre-set range, for example being higher than 0.9 or being equal to 1. At 0, an operating state "no cavitation" can optionally be used. This value assignment is a matter of preference, it is only necessary for the correct operating states 511 to be assigned to the cavitation scores 411 as required. For example, each cavitation score 411 could also correspond to a more specific cavitation risk level.

[0069] Step 600 is executed after the operating states are determined. Depending on the operating state 511, The pressure will thus be regulated by the pressure controller 91. If the operating state "low cavitation risk" 511a is detected, the control signal to decrease the back-pressure 611c will thus be sent by the control device 9 to the pressure controller 91 via the communication link 92 and the pressure controller will cause a reduction in the back pressure, reducing the energy consumption of the centrifuge. If the operating state "high cavitation risk" is detected, in this exemplary embedment the control signal 611 to maintain the back-pressure 611a will be sent to the pressure controller 91. Should the operating state" cavitation imminent" be determined, the pressure controller 91 will receive a signal to increase the back pressure 611b, to prevent the cavitation from occurring. An operating state "cavitation" means that cavitation has occurred, and in this case an alert will be sent 611d, so the operation of the centrifuge might be ended. In the other cases the operation resumes, as indicated in Fig. 2b.

[0070] Fig. 3 outlines the operation of the control device 9, as well as the signal flow in more detail. It is shown that structure-borne sound data 111, which is a temporal course of a structure born sound signal (as described above), is collected by the structure-borne sound sensor 22 on the centrifuge as part of a sampling of the signal data 101. This data is transmitted to the control device 9 and converted by this into a set of discrete frequency spectrum data 211. It is possible that the sensor 22 provided digital or analogue data. In the latter case the control device 9 should comprise an analogue-to-digital converter (ADC) to interpret the structure-borne sound signal. In an alternative embodiment that is not shown here, the structure-borne sound sensor carries out the conversion of the time-domain signal into the frequency domain as part of its operation and provided this as the frequency spectrum data 211 to the control device 9.

[0071] The control device 9 comprises the machine learning tool 301, either as a computer program loaded onto a storage unit or as a hardware circuit, although the former implementation is preferential as it is cheaper. The control device preferably comprises a processor, microprocessor or microcontroller, but may also include an AI accelerator device for faster implementation of the machine learning tool's algorithm 3011, especially for faster training.

[0072] The frequency spectrum data 211 is input into the machine learning tool 301 and evaluated by its algorithm 3011, which outputs the cavitation score 411 and provides this to the control device 9 for the determination of the operating state 511 as described above. Depending on the operating state, a control signal 611 is sent to the pressure controller 91, which is configured to regulate the back-pressure as described for Fig. 2b.

[0073] The machine learning tool 301 in this embodiment was trained through a method according to the present invention using training data 101', which is shown in Fig. 4.

[0074] The training data 101' comprises a multitude of sets of training frequency spectrum data 211', acquired by transforming a multitude of temporal courses of training structure-borne sound signals 111' acquired on the centrifuge, or another centrifuge of essentially identical build, into the frequency domain, as an input and respective training (i.e. label) cavitation scores 411' as an output.

[0075] The training structure-borne sound signals 111' can have been detected using the structure-borne sound sensor 22 on the centrifuge or one of equivalent build, while the label cavitation scores 411' could be ones detected using a more sensitive cavitation detection device.

[0076] The training of the machine learning tool 301 of Fig. 3, using supervised learning, is conducted using typical methods according to the state of the art. In particular, the training data 101' is input into the, at his point before the tool is trained, yet untrained machine learning tool 301'. Here the untrained machine learning tool 301' receives the training frequency spectrum data 211', and evaluates it using an iterative optimisation algorithm or a generative training algorithm, yielding output cavitation scores 411 which the machine learning tool then compares to the training (i.e. label) cavitation scores 411' that are defined as part of the training data 101' in training (comparison) step 401'. According to this result, the machine learning tool 301' adjusts and evaluates its function in step 400" if it identifies that the algorithm does not yield the output cavitation score 411' defined by the training data 101' for a particular input 211', to improve the fit. This process including step 400" is then repeated iteratively until a preset maximum error margin is achieved when a set of training frequency spectrum data 211' is evaluated by the untrained computer-implemented algorithm 3011' to yield a cavitation score 411, meaning that this cavitation score 411 should be within a certain maximum range of the cavitation score 411' assigned to the dataset 101', i.e. the label 411' of the set of frequency spectrum data 211'.

[0077] This process is then preferably repeated for the validation data, i.e. the sets of validation frequency spectrum data are input into the machine learning-tool, and the algorithm's parameters are then iteratively adjusted until the algorithm 3011' also yields the correct output validation cavitation score for a respective input of a set of validation frequency spectrum data within a preferred error margin. The performance can be evaluated using test data, without step 400" and with step 400' instead only comprising an evaluation step that outputs the error margin to a user.List of reference signs

[0078] 1Bowl 2Drive spindle 3Sensor 4Product inlet 5Solids discharge openings 6Sliding piston 7Solids catcher 8Solids holding space 9Control device 91Pressure controller 92Communication link 10Bowl bottom 11Bowl top 12Hood 13Outlet 14Disc stack 15Line for control water 16Valve arrangement 17Solenoid valve block 18Communication link 19Supply line 20Sensor 21Communication link 22Structure-borne sound sensor 23Communication link 40Inlet pipe 100 - 600Method steps 400'Training step 400"Adjustment and Evaluation 101Signal data 101'Training data 111Structure-borne sound data 111'Training structure-borne sound data 211Frequency spectrum data 211'Training frequency spectrum data 301Machine learning tool 301'Machine learning tool under training 3011Algorithm 3011'Algorithm under training 411Cavitation score 411'Label cavitation score 511Operating state 511aOperating state "low cavitation risk" 511bOperating state "high cavitation risk" 511cOperating state "cavitation imminent" 511dOperating state "cavitation" 611Control signal 611aMaintain back-pressure 611bIncrease back-pressure 611cDecrease back-pressure 611dSend alert LClear phase PStarting product SSolid matter DAxis of rotation

Claims

1. Method for operating a centrifuge that has at least one rotatable bowl (1) with a pressure controller (91), configured to set a level of a back-pressure inside the bowl (1), and that has an inlet pipe (40) and that has at least one liquid outlet (13), with which, during operation, a flowable starting product (P) is separated into different liquid and / or solids phases in the centrifugal field of the rotating bowl (1), the method comprising at least the following computer-implemented steps for recognising an operating state (511): 100Sampling signal data (101) with at least one sensor (22) arranged on the centrifuge, the signal data (101) comprising a discrete-time structure-borne sound signal (111) representing a temporal course of a structure-borne sound signal detected using at least one structure-borne sound sensor (22);200Transforming the discrete-time structure-borne sound signal (111) to the frequency domain, to obtain a set of frequency spectrum data (211);300Inputting the set of frequency spectrum data (211) into a computer-implemented machine learning tool (301) executed on a control device (9);400Evaluating the set of frequency spectrum data (211) through a computer-implemented algorithm (3011) of the machine learning tool (301), hence determining a cavitation score (411), and providing the cavitation score (411) to the control device (9) of the centrifuge, by the computer-implemented algorithm (3011) of the machine leaning tool (301);500Determining of an operating state (511) of the centrifuge based on the cavitation score (411) by the control device (9);600Regulating the back-pressure inside the bowl (1) of the centrifuge through the control device (9) by sending a control signal (611) that is dependent on the operating state (511) from the control device (9) to the pressure controller (91).

2. Method according to claim 1, wherein the structure-borne sound signals (111) are acquired over a frequency range from 0 kHz to 30 kHz, preferably from 10 kHz to 20 kHz and / or up to 2 kHz.

3. Method according to claim 1 or 2, wherein the machine learning tool (301) was trained using training data (101') that comprises a multitude of sets of training frequency spectrum data (211'), acquired by transforming a multitude of temporal courses of training structure-borne sound signals (111') acquired on the centrifuge, or another centrifuge of essentially identical build, into the frequency domain, as an input and respective training cavitation scores (411') as an output.

4. Method according to claim 3, wherein the training structure-borne sound signals (111') are acquired over a frequency range of up to 10 kHz, up to 20 kHz, up to 1 or more MHz, and most preferably from 10 kHz to 20 kHz.

5. Method according to any one of the claims 3 or 4, wherein the training cavitation scores (411') correspond to label cavitation scores as an answer key.

6. Method according to any one of claims 3 to 5, wherein the training structure-borne sound signals (111') are acquired by the at least one structure-borne sound sensor (22) and the training cavitation scores (411') are ones detected simultaneously using a cavitation detection device that detects structure-borne sound signals up to a higher frequency than the structure-borne sound sensor (22).

7. Method according to any one of claims 3 to 6, wherein the machine learning tool (301) was tuned using validation data that comprises a multitude of sets of validation frequency spectrum data, acquired by transforming a multitude of temporal courses of validation structure-borne sound signals acquired on the centrifuge, or another centrifuge of essentially identical build, into the frequency domain, as an input and respective validation cavitation scores as an output.

8. Method according to claim 7, wherein the validation structure-borne sound signals are acquired over a frequency range of up to 10 kHz, up to 20 kHz, from 10 kHz to 20 kHz, or up to 1 or more MHz.

9. Method according to one of claims 3 to 8, wherein the machine learning tool (301) was verified using test data that comprises a multitude of sets of test frequency spectrum data, acquired by transforming a multitude of temporal courses of test structure-borne sound signals acquired on the centrifuge, or another centrifuge of essentially identical build, into the frequency domain, as an input and respective test cavitation scores as an output.

10. Method according to claim 9, wherein the test structure-borne sound signals are acquired over a frequency range of up to 10 kHz, up to 20 kHz, from 10 kHz to 20 kHz, or up to 1 or more MHz.

11. Method according to any one of the preceding claims, wherein a cavitation score (411) within a first pre-set value range corresponds to an operating state "cavitation imminent" (511c) that indicates that an occurrence of a cavitation inside the bowl (1) is imminent, and wherein a cavitation score (411) within a second pre-set value range corresponds to an operating state "low cavitation risk" (511a) that indicates that an occurrence of a cavitation inside the bowl (1) is unlikely.

12. Method according to claim 11, wherein the control device (9) is configured to increase the back-pressure inside the bowl of the centrifuge when the operating state "cavitation imminent" (511c) or an operating state "high cavitation risk" (511b) is detected, by sending a control signal (611b) to increase the back pressure inside the bowl (1) to the pressure controller (91), and wherein the control device (9) is configured to decrease the back-pressure inside the bowl (1) of the centrifuge when the operating state "no cavitation" or "low cavitation risk" (511a) is detected, by sending a control signal (611c) to decrease the back pressure inside the bowl to the pressure controller.

13. Method according to any one of the preceding claims, wherein the pressure controller (91) is configured to increase or decrease the back-pressure inside the bowl (1) depending on the control signal (611) from the control device (9).

14. Method according to any one of the preceding claims, wherein each of the structure-borne sound signals is acquired using the at least one structure-borne sound sensor (22), or at least one essentially identical structure-borne sound sensor on another centrifuge of essentially identical build.

15. Method according to any one of the preceding claims, wherein the signal data (101) comprises temperature data, humidity data, process data, in particular bowl speed or processed fluid, or pressure, in particular back-pressure, data.

16. Method according to any one of the preceding claims, wherein the control device (9) is configured to issue a warning message when the operating state "cavitation imminent" (511c) is detected.

17. Method according to claim 16, wherein the computer-implemented machine learning tool (301) has a classification algorithm (3011), preferably based on an artificial neural network, nearest neighbours, linear regression, naïve Bayes, random forest, a decision tree, a vector machine, or an autoencoder.

18. Computer-implemented method for training a machine learning tool (301) that provides a cavitation score (411) based on a structure-borne sound signal (111) detected on a centrifuge, the method including a step of training the machine learning tool in an untrained state (301') using training data (101 ') that comprises a multitude of sets of training frequency spectrum data (211'), acquired by transforming a multitude of temporal courses of training structure-borne sound signals (111') acquired on the centrifuge, or another centrifuge of essentially identical build, into the frequency domain, as an input and respective training cavitation scores (411') as an output.

19. Computer-implemented method according to claim 18, wherein the method includes a step of tuning the machine learning tool (301) using validation data that comprises a multitude of sets of validation frequency spectrum data, acquired by transforming a multitude of temporal courses of validation structure-borne sound signals acquired on the centrifuge, or another centrifuge of essentially identical build, into the frequency domain, as an input and respective validation cavitation scores as an output.

20. Computer program including a set of instructions that, when executed on a control device (9), cause the control device (9) to implement the method according to any one of claims 1 to 19.

21. Centrifuge, to which a structure-borne sound sensor is attached and which has at least one rotatable bowl with an inlet pipe and at least one liquid outlet and / or one solids outlet, with which a flowable starting product can be separated into different liquid and / or solid phases in the centrifugal field of the rotating bowl during operation, wherein the centrifuge further comprises a control device which can detect a structure-borne sound signal of the centrifuge by means of the structure-borne sound sensor and which comprises a processing device which implements the method for operating the centrifuge according to any one of claims 1 to 17.

22. Centrifuge according to claim 21, characterised in that the control device is configured to control the back pressure according to claim 11 or 12.

23. Centrifuge according to claim 21 or 22, characterised in that the processing device comprises an artificial intelligence accelerator that accelerates the implementation of the machine learning algorithm.

24. Centrifuge according to any one of claims 21 to 23, characterised in that the structure-borne sound sensor is a vibration sensor, preferably an accelerometer.

Images