A method for determining an operating condition of a reciprocating compressor

Abstract

A computer-implemented method for determining an operating condition of a reciprocating compressor, the method comprising the steps of: a. receiving (100) time-series sensory data (1) of the reciprocating compressor; b. creating (101) a dataset (3) comprising at least one of: a time derivative of the time-series sensory data (2) as a function of the time-series sensory data (1); and time-delay embeddings of the time-series sensory data; c. extracting (102) a set of data features from the dataset (3); d. performing (103) an analysis on the set of extracted data features to determine and operating condition of the reciprocating compressor.

Description

[0001] A method for determining an operating condition of a reciprocating compressor

[0002] The present invention relates to the field of condition monitoring of reciprocating compressors . In particular, it relates to a computer-implemented method for determining an operating condition of a reciprocating compressor ; to a computer-implemented method for determining at least one cause of an emerging malfunction or a mal function of a reciprocating compressor ; to a data processing system; to a reciprocating compressor ; to a computer program; and to a computer-readable medium .

[0003] Reciprocating compressors , also known as piston compressors , are mechanical devices that generate compressed air or gas by using pistons driven by a crankshaft in a cyclic motion . This type of compressor is widely used in various industries , including refrigeration, oil refineries , gas pipelines , and natural gas processing plants , where high-pressure gas is required . The process involves drawing in air or gas through an intake valve , compressing it in a confined space , and then releasing it through an exhaust valve . Reciprocating compressors are prone to wear and tear . Therefore , they require considerable maintenance , and the condition of the compressor must be assessed and / or monitored regularly to ensure operational reliability and to prevent failures .

[0004] One common method for condition monitoring of reciprocating compressors is vibration analysis , where sensors are placed on the compressor to detect unusual vibrations that can indicate mechanical issues such as misalignment , imbalance , or bearing failures . Temperature monitoring is also commonly used, as excessive heat can be a sign of inadequate lubrication, cooling system failures , or excessive friction . Additionally, acoustic emissions can be monitored; these are sounds produced by the compressor that can reveal problems like valve leaks or piston slap, for example .

[0005] In EP 2646783 Bl ( JOHN CRANE UK LIMITED, 30 June 2020 ) , condition monitoring of reciprocating equipment is achieved using an acoustic emission detection system that records time domain data among other data . This method captures variations in signal intensity over time , speci fically focusing on data from a single rotation cycle to detect potential defects .

[0006] However, a key disadvantage of this time domain analysis approach is the di f ficulty in accurately extracting data for j ust one cycle of the reciprocating equipment . Misidenti fications in the data segmentation can occur due to signal overlap, operational noise , or speed variability, potentially leading to erroneous diagnostics . Such inaccuracies can cause false positives (unnecessary repairs ) or false negatives , where actual defects are missed, making the method less reliable in environments with fluctuating signal characteristics .

[0007] As another example , the condition monitoring of a reciprocating compressor by measuring the pressure as a function of the crank angle and detecting changes in the data over time is shown in EP 3786495 Al (BURCKHARDT COMPRESS ION AG, 03 March 2021 ) .

[0008] However, this approach brings with it a number of challenges that can make the process imprecise and time-consuming . Firstly, the collected data must be synchronised, which requires a so called trigger signal . I f the trigger signal is of f , the data collected could be inaccurate , which af fects the subsequent analysis . This is especially problematic in multi-stage compressors . Secondly, also the crank angle value has to be measured accurately and it is dependent on a reference value that must be set manually . This impedes the scalability of such a process .

[0009] In another example , US 8958995 B2 (HONEYWELL INTERNATIONAL INC, 17 February 2015 ) discloses a method that uses the pressure measurement as a function of the varying cylinder volume to create a pressure-volume diagram and compare the pressure-volume diagram to a reference pressure-volume diagram to detect a malfunction .

[0010] The inventors of the present invention have found that one of the disadvantages of such a method is that it requires the precise determination of the cylinder volume . However, the cylinder volume , which changes during a crank shaft revolution, must be estimated in a rather laborious manner by approximating a multitude of geometrical values . This is imprecise because each geometrical value entails a certain degree of measurement inaccuracy and is also susceptible to user errors . Apart from this , it is very labour-intensive , as the cylinder volumes di f fer from compressor to compressor and these settings must therefore be made individually for each compressor .

[0011] It is therefore generally disadvantageous i f the data used for condition monitoring is dependent on values that on the one hand are subj ect to errors and on the other hand are resource-intensive to determine .

[0012] It is therefore an obj ect of the present invention to overcome at least some , i f not all shortcomings of the prior art . In particular, it is an obj ect of the present invention to provide an improved method for condition monitoring of reciprocating compressors that is more reliable , more accurate and less prone to errors and mistakes than the prior art methods . The problem is solved by a computer-implemented method for determining an operating condition of a reciprocating compressor ; by a computer-implemented method for determining at least one cause of an emerging mal function or a mal function of a reciprocating compressor ; by a data processing system; by a reciprocating compressor ; by a computer program; and by a computer-readable medium . Advantageous embodiments and further developments are the subj ect of the dependent claims .

[0013] A first aspect of the invention relates to a computer-implemented method for determining an operating condition of a reciprocating compressor . The method comprises the following steps :

[0014] In step a, time-series sensory data of the reciprocating compressor are received . Typically, the method of the first aspect of the invention is carried out in a data processing system in which the data are received from at least one sensor .

[0015] In step b, a dataset is created which comprises at least one of the following : a time derivative of the time-series sensory data as a function of the time-series sensory data ; and time-delay embeddings of the time-series sensory data .

[0016] In step c, a set of data features from the dataset is extracted .

[0017] In step d . , an analysis is performed on the set of extracted data features to determine and operating condition of the reciprocating compressor .

[0018] In the context of the present invention, an "operating condition" refers to the speci fic set of parameters under which the compressor is functioning at any given time . This can include factors such as pressure , temperature , flow rate , and mechanical speed . A "normal operating condition" is characteri zed by the compressor running within its designed speci fications and ef ficiency parameters , ensuring optimal performance without any signs of stress or anomaly . An "operating condition with an emerging mal function" indicates the initial stages where deviations from normal performance begin to appear, potentially due to issues like wear, inadequate lubrication, or minor mechanical failures . These deviations can be subtle and may not immediately impact the overall performance but signal the onset of potential problems . Lastly, an "operating condition with a mal function" is when the compressor operates outside of its normal speci fications due to signi ficant mechanical or operational failures , leading to reduced ef ficiency, potential damage , or complete breakdown . Each condition requires di f ferent levels of monitoring, maintenance , and intervention to ensure the longevity and ef ficiency of the compressor . It is understood that the transition between "normal operating condition" , "operating condition with an emerging mal function" and "operating condition with a mal function" is gradual . The individual conditions can be distinguished from each other, for example , by threshold values .

[0019] An operating condition of the reciprocating compressor can also refer to a speci fic component of the reciprocating compressor, for example to the operating condition of a speci fic valve or a crankshaft . In particular, the operating condition of speci fic components to be determined are components related to the compressor cylinder, such as the piston, cylinder liner, valves , cylinder head, rod and crankshaft bearings and gaskets .

[0020] In the context of the present invention, time-series sensory data refers to a sequence of data points collected over time through sensory inputs . The data can be directly measured, consisting of raw, unprocessed outputs from sensors that monitor variables such as temperature or pressure . Additionally, timeseries sensory data can also be derived data, where the raw measurements undergo trans formations or calculations to generate new data points that of fer more insights or are more suited for speci fic analyses .

[0021] It can be a single time-series sensory data type either from a single sensor or from multiple sensors of the same type . However, it can also be several time-series sensory data types from several sensors of di f ferent types .

[0022] Preferably, the sampling rate of the time-series sensory data ranges from 1 kHz to 100 kHz , more preferably ranges from 5 kHz to 50 kHz , most preferably ranges from 10 kHz to 30 kHz .

[0023] The sensor data of the reciprocating compressor is data measured in or on the compressor, or at a distance from the compressor ( e . g . acoustic data ) . The data can be used to determine an operating condition of the compressor or a part of it .

[0024] In step b . of the method according to the invention, a dataset is created from the time-series sensory data, which can comprise a time derivative of the time-series sensory data as a function of the time-series sensory data . A graphical representation of the time derivative of the sensory data on the y-axis and the sensory data on the x-axis would correspond to an orbit plot , wherein one shaft revolution corresponds to one graphically represented revolution .

[0025] Additionally or alternatively, the dataset created in step b . comprises time-delay embeddings of the time-series sensory data . Time-delay embeddings (for example Takens' embeddings) are created by shifting the time-series sensory data by one or more time steps. For example, if the original time-series sensory data are cylinder pressure data and the compressor speed of the measured data is one revolution per second, a time delay of one second would create a new column in the dataset where each entry is the previous revolution's cylinder pressure. The created dataset can include more than one time-delay embedding, and the time delays can be, for example, in the order of 0.01 ms, 0.1 ms, 1 ms, 1 s, or 1 min. For further information on the subject of time-delay embeddings, reference is made to document RAVINSHANKER, N. et al. (An introduction to persistent homology for time series. WIRFs Comput . Stat. 2021. 13:el548) , which also covers Takens' delay embedding for time series (Section 2.3.1) .

[0026] The dataset created in step b. may exist in numerical form and / or it may be represented graphically, for example to be displayed to a user via a user interface.

[0027] In the context of the present invention, the term "or" is to be understood as an exclusionary disjunction. "A or B" states that exactly one of the two statements is true (if the disjunction is true) .

[0028] In step c. data features are extracted from the dataset created in step b.. Data feature extraction typically involves selecting or transforming raw data into a set of variables, known as data features, that are useful for direct analysis or predictive modelling .

[0029] In the case of time derivative sensory data as a function of the sensory data, there is a multitude of possible data features that can be extracted from the dataset. For example, slopes of different stages of a crankshaft revolution (e.g. compression phase, expansion phase, ...) can be extracted; maximum and / or minimum values can be extracted; and integrals within certain value ranges can be determined (e.g. the area within the orbit plot) .

[0030] Feature extraction from time-delay embeddings can be performed according to the instructions in Section 4 of RAVINSHANKER, N. et al. (An introduction to persistent homology for time series. WIRFs Comput . Stat. 2021. 13:el548) .

[0031] It is possible that only one data feature is extracted in step c., but preferably multiple data features are extracted.

[0032] Preferably, the data features in step c. are extracted for a single shaft revolution, and / or the mean value of the extracted data features of a certain time period (e.g. one minute or one hour) is used for the analysis in step d. .

[0033] In step d., the set of extracted data features is preferably analysed using an automated data analysis method, which include the following methods: threshold-based analysis works by comparing data features against predefined normal ranges; deviations indicate potential issues. Trend analysis monitors data over time, with gradual deviations suggesting emerging problems and sharp changes indicating malfunctions. Statistical analysis compares data features to historical norms (e.g. mean values and standard deviations) , where anomalies point to potential malfunctions. Pattern recognition and machine learning classify data based on learned patterns, identifying normal, emerging, and malfunctioning states. Signal processing and frequency analysis examine the frequency components of vibrations or acoustic signals, where shifts or new frequencies can indicate issues. Correlation and multivariate analysis look for expected relationships among multiple features . The methods can also be combined with each other .

[0034] In addition or as an alternative to an automated analysis , the extracted data features may be analysed such that they are presented to a user in numerical and / or graphical form; and that a user input is received as a result of an analysis of the data features performed by the user . In addition to the extracted data features , the time-series sensory data may also be presented to the user in numerical and / or graphical form .

[0035] For example , the analysis in step d . comprises a step of comparing the data features extracted in step c . to reference values . Preferably, reference values are simulated or measured desired values representing a compressor with a normal operating condition; and / or reference values are values measured previously on the compressor to be assessed in the method according to the invention . Alternatively, the extracted data features could be used without reference values , e . g . by applying statistical analysis on the data features .

[0036] The method according to the invention works with both singleacting and double-acting compressors , as well as with both single-stage and multi-stage compressors .

[0037] The method of the present invention allows a precise determination of an operating condition of a reciprocating compressor . The method according to the invention solves problems of the prior art , in particular, i f the dataset created in step b . contains a time derivative of the times-series sensory data as a function of the time-series sensory data . For example, the present method eliminates the need to determine a start point and an end point of a crank rotation, as is the case, for example, with known time-domain analysis approaches (e.g. EP 2646783 Bl, JOHN CRANE UK LIMITED, 30 June 2020) . The influence of faulty data segmentation, for example due to noise signals, is therefore reduced, if not prevented. The present method is also less error-prone when the measured data (e.g. pressure data) is highly variable, i.e. when the compressor speed changes rapidly or fluctuates greatly. For example, the time derivative time-series data changes to the same extent as the corresponding time-series data when the compressor speed changes. This means that the time derivative sensory data as a function of the associated sensory data is decoupled from the compressor speed.

[0038] In addition, the method according to the invention only requires the time-series sensory data without additional measured values or parameters. Such additional measured values or parameters, which in turn may be subject to errors and / or time-consuming to determine, have already been described above. These include, for example, the determination of the crank angle in methods in which a measured value (for example the pressure) is used as a function of the crank angle for condition mounting. In the present method, for example, the crank angle does not need to be determined. Nor does the variable cylinder volume need to be determined, which would be based on an inaccurate estimate. In other words, the condition monitoring in this invention is based on intrinsic analysis of time-series sensory data. In contrast, prior art methods rely on additional measured values and / or parameters for extrinsic analysis for the condition monitoring of a compressor. This prevents or at least mitigates errors and (user) mistakes; and it saves considerable amounts of time when commissioning a compressor's condition monitoring system, as there is no need for manual set-up and manual adj ustment of machine-speci fic values .

[0039] Overall , the method according to the invention reduces false positives and false negatives , which reduces or prevents unnecessary repairs and nevertheless detects faulty compressor states in the form of mal functions or emerging mal functions . The method according to the invention thus provides a solution for condition monitoring of reciprocating compressors that is particularly reliable , accurate and less prone to errors and mistakes . Furthermore , the method according to the invention enables improved scalability of condition monitoring of reciprocating compressors .

[0040] In a preferred embodiment of the method of the first aspect of the invention, the time-series sensory data exhibit cyclic variations depending on a revolution of a crankshaft of the reciprocating compressor . The cyclic variations of the time-series sensory data ( such as the cylinder pressure ) in a reciprocating compressor are due to the periodic motion of the pistons within the cylinder . It was found that the method according to the invention works exceptionally well for time-series sensory data with such cyclic variations .

[0041] In a further preferred embodiment of the first aspect , the timeseries sensory data are selected from the group consisting of pressure data, temperature data, vibrational data, acoustic data, strain data, position data, proximity sensor data, image data, and video data ; preferably selected from the group consisting of pressure data, temperature data, and acoustic data .

[0042] Preferably, the time-series sensory data is measured within a compression space of the reciprocating compressor, in particular pressure data and / or temperature data are measured within a compression space of the reciprocating compressor. Compression space includes all volumes of a reciprocating compressor that have a different pressure to the ambient pressure during compressor operation. Furthermore, the pressure in the compression space also changes relative to the suction and discharge pressure during a compression cycle.

[0043] Sensors for measuring pressure data in a reciprocating compressor typically measure at least one of the suction (inlet) pressure, the discharge (outlet) pressure, the interstage pressure, the crankcase pressure, and the cylinder pressure. Temperature data are typically obtained as at least one of suction (inlet) temperature, discharge (outlet) temperature, inter-stage temperature, cylinder head temperature, crankcase temperature, and coolant temperature.

[0044] Vibrational data measured include displacement (the distance a point on a machine moves from its resting point) , velocity (speed of this movement) , and acceleration (rate of change of this movement) . Vibrational data can be received as spectral data and transformed into time-series data. Phase data can also be included in the vibrational data (e.g. as the phase difference over time) to measure the phase angle difference between different vibration signals.

[0045] Acoustic data refers to sound information in a broad sense. In essence, it comprises sound pressure levels (SPL) , but it may include other information such as the spatial distribution of the sound intensity at different locations of the reciprocating compressor, and / or the directionality of the noise. Acoustic data may comprise data collected from an array of multiple acoustic sensors. Strain data, used for detecting mechanical stresses and deformations occurring within the compressor components , for example , may comprise at least one of shear strain data, bending strain data, tensile strain data, and compressive strain data .

[0046] The position data can be , for example , the position data of a piston rod, a crankshaft , or a valve of the compressor . A proximity sensor can detect the presence , absence , and distance without physical contact . Proximity sensor data may be used to detect the piston rod position, a position of a valve , the crankshaft position, crosshead position, the position of the clearance pocket adj usters , and the position of an actuator . Inductive , capacitive , optical , magnetic and ultrasonic proximity sensors are known in the art , for example .

[0047] Image data and / or video data received as time-series sensory data may comprise the image and / or video data themselves . However, it also may include information extracted from this image and / or video data . This extracted information can encompass metrics and features such as movement patterns , obj ect recognition, colour histograms , and edge detection results . For instance , in the context of condition monitoring for reciprocating compressors , video data might be analysed to detect anomalies in the motion of the pistons , measure the vibration frequencies , or identi fy changes in the appearance of critical components that could indicate wear or damage . Additionally, thermal imaging data could provide temperature distribution maps that reveal hotspots and potential overheating issues . The data recorded or extracted over time can be received as a time series .

[0048] Time-series sensory data can also include correlation data between multiple di f ferent data types . For example , acoustic data can be correlated with vibration data which may help confirm the presence of faults and identi fy their location .

[0049] The data types mentioned in this embodiment , in particular pressure data, temperature data, and acoustic data, are particularly suitable as time-series sensory data in the method according to the invention .

[0050] In another preferred embodiment of the first aspect of the invention, the time-series sensory data comprise or are pressure data, preferably pressure data selected from the group consisting of suction pressure data, discharge pressure data, cylinder pressure data, interstage pressure data, and crankcase pressure data, more preferably the pressure data are cylinder pressure data .

[0051] In a further preferred embodiment , the time-series sensory data do not comprise at least one sensory data type from the group consisting of displacement data, crank angle data and volume data .

[0052] In the context of the present invention, displacement data refers to data that describes the displacement of a moving part in the compressor ; for example , displacement data of the piston rod or the connecting rod .

[0053] As described above , the inventors have found that synchronising and / or acquiring the crank angle data and the volume data ( especially the cylinder volume ) is time-consuming and error-prone , which compromises the scalability and the accuracy of the method . The method according to the invention does not require these values . In another preferred embodiment of the first aspect of the invention, the dataset created in step b . comprises a time derivative of the time-series sensory data as a function of the timeseries sensory data . Each data feature extracted in step c . is of any of the following data feature types :

[0054] - the slope and / or the y-intercept of a first linear approximation of a first set of data points of the dataset , preferably a linear approximation of a set of data points of the dataset representing at least a part of the compression phase of a compression cycle ;

[0055] - the slope and / or the y-intercept of a second linear approximation of a second set of data points of the dataset , preferably a linear approximation of a set of data points of the dataset representing at least a part of the expansion phase of a compression cycle ;

[0056] - the maximum time derivative data point and, optionally, the corresponding time-series data point ;

[0057] - a first integral value of a third set of data points of the dataset , preferably an integral value of a set of data points of the dataset representing at least a part of the compression phase of a compression cycle ;

[0058] - a second integral value of a fourth set of data points of the dataset , preferably an integral value of a set of data points of the dataset representing at least a part of the expansion phase of a compression cycle ;

[0059] - the di f ference between the first integral value and the second integral value .

[0060] In the context of the present invention, a compression cycle refers to the sequence of operations that a reciprocating compressor undergoes to compress a gas within its cylinders . For a single-acting compressor, it typically includes a suction phase , where the intake valve opens to allow the gas to enter the cylinder as the piston moves downward; the compression phase , where the intake valve closes and the piston moves upward, compressing the gas ; the discharge phase where the exhaust valve opens to release the compressed gas as the piston reaches the top of its stroke ; and the expansion phase , where the exhaust valve closes , and the piston moves downward again, preparing for the next intake .

[0061] With respect to the first two data feature types , it was found that in the case of the dataset of step b . comprising a time derivative of the time-series sensory data as a function of the time-series sensory data, the data, in particular cylinder pressure data, have data ranges that represent an approximate linear relationship . This can be visualised by displaying the data as an orbit plot . One such data range can, for example , represent the compression phase of the compressor, i . e . the data range that is recorded during the compression of the gas . Another data range with an approximately linear data relationship can represent the expansion phase of the cylinder . The data ranges from which the slope and / or the y-intercept are derived are selected in such a way that the respective end points lie in or at least at the end positions of the linear relationships .

[0062] Linear approximations are computationally particularly ef ficient ( e . g . compared to polynomial regression) and robust against noise , making them reliable for continuous monitoring and early fault detection . The possibility of deriving linear relationships from the time-series sensory data, which typically do not have a linear relationship, as a result of the method according to the invention, is therefore particularly advantageous . The maximum time derivative data point corresponds to the point at which the strongest positive change in the measured sensory data is measured, which typically occurs during or towards the end of the piston stroke . It has been found that changes in this maximum point indicate mal functions and emerging mal functions particularly early and accurately .

[0063] For the three data feature types relating to integrals , the limits of integration are preferably selected so that the data ranges to be integrated are not multivalued . This means that i f the time derivative data were plotted on the y-axis and the sensor data itsel f on the x-axis , the data range to be integrated is selected so that there is no more than one y-value for each x-value . I f the time-series sensory data are pressure data of a speci fic compressor stage , these limits of integration can be , for example , the suction pressure and the discharge pressure of this compressor stage .

[0064] Since the time derivative data as a function of the sensory data of a compressor can typically be represented by an orbit plot due to their cyclic nature , in this case the di f ference between an integral of the data values representing the compression phase of the compression cycle and an integral of the data values representing the expansion phase of the compression cycle corresponds to the area within the orbit plot , at least approximately .

[0065] Changes in these integral values are also decisive for the detection and early detection of mal functions , as found by the inventors . In another preferred embodiment of the first aspect , the timeseries sensory data are pressure data, preferably cylinder pressure data . The dataset created in step b . comprises the pressure derivative as a function of the pressure data . The data features extracted in step c . are at least one of , preferably at least the maj ority of , more preferably all but one of the following data features :

[0066] - the slope and / or the y-intercept of a linear approximation of the pressure derivative as a function of the pressure data representing at least a part of the compression phase of a compression cycle ;

[0067] - the slope and / or the y-intercept of a linear approximation of the pressure derivative as a function of the pressure data representing at least a part of the expansion phase of a compression cycle ;

[0068] - the maximum pressure derivative value and, optionally, the corresponding pressure value ;

[0069] - a first integral value of the pressure derivative as a function of the pressure data representing at least a part of the compression phase of a compression cycle ;

[0070] - a second integral value of the pressure derivative as a function of the pressure data representing at least a part of the expansion phase of a compression cycle ;

[0071] - the di f ference between the first integral value and the second integral value .

[0072] In the context of the present invention, the maj ority of a number of entities corresponds to more than 50 % of the number of entities .

[0073] In a further preferred embodiment of the first aspect , the analysis performed in step d . comprises a user analysis . The user analysis comprises the steps of : - displaying the extracted data features and / or the received time-series sensory data to a user on a user interface ;

[0074] - optionally, displaying reference values to the user on the user interface ;

[0075] - receiving a user input .

[0076] The extracted data features and / or the received time-series sensory data can be displayed to the user in numerical form and / or as a data plot . The user is thus prompted to assess the data displayed and to report any deviations or discrepancies in the form of user input . Reference values can also be displayed . These can be calculated or measured normal values to determine deviations from these normal values ; or they can be historical values of the same sensor data to analyse the development of the data over time . The user input contains information about the operating condition . For example , the user input may include an indication of whether the user considers the operating condition of the compressor to be a normal operating condition, an operating condition with an emerging mal function, or an operating condition with a mal function . The user input received in this way can be used alone or in addition to , for example , an automated analysis to determine an operating condition of the compressor .

[0077] In another preferred embodiment , the analysis performed in step d . comprises a computational analysis , and performing the computational analysis on the extracted data features comprises applying a trained machine-learning model on the extracted data features . Optionally, the machine-learning model is trained using an unsupervised, a semi-supervised, or a supervised learning algorithm . Preferably, the algorithm is at least one of an ensemble learning algorithm and a neural network algorithm . Preferably, the ensemble learning algorithm is at least one of a boosting algorithm (preferably a gradient boosting algorithm) and a random forest algorithm .

[0078] In a further preferred embodiment of the first aspect , the analysis performed in step d . comprises a computational analysis , and performing the computational analysis on the extracted data features comprises applying a weighted average formula on the extracted data features .

[0079] In a preferred embodiment of the method according to the first aspect , the analysis performed in step d . comprises a combined user analysis and a computational analysis .

[0080] In a further preferred embodiment , the operating condition is selected from the group consisting of : a normal operating condition; an operating condition with an emerging mal function; and an operating condition with a mal function . These types of operating conditions have been described above .

[0081] In another preferred embodiment , the operating condition is determined by determining ( a ) deviation ( s ) between the extracted data features and reference values . As described above , reference values can be standard values representing a normal operating condition and / or historical values of the data features . Historical values can represent the operating condition of the reciprocating compressor of one or more points in time .

[0082] In a further preferred embodiment of the first aspect , the method comprises at least one additional step selected from the group consisting of :

[0083] - noti fying a user of the determined operating condition and / or ( a ) determined deviation ( s ) ; - prompting a user to take measures to restore a normal operating condition and / or to stop at least part of the reciprocating compressor ; and

[0084] - sending instructions to a control system of the reciprocating compressor to stop at least part of the reciprocating compressor .

[0085] Any ( one ) of these steps is / are executed i f the determined operating condition does not correspond to the normal operating condition .

[0086] An operating condition with an emerging mal function or with a mal function typically triggers an operator to perform, for example , visual inspection and / or maintenance of the reciprocating compressor . This includes determining the cause of the abnormal operating condition, replacing compressor components , and informing the system that there is no problem and that the current condition is the new 'normal operating condition' .

[0087] In a further preferred embodiment , the method comprises , preferably as part of step d . , an additional step of determining a cause for the operating condition, i f the operating condition is not a normal operating condition .

[0088] In another preferred embodiment of the first aspect , the method comprises , before step a . , an additional step of acquiring timeseries sensory data using at least one sensor, preferably selected from the group consisting of pressure sensors , temperature sensors , vibrational sensors , and acoustic sensors . Other possible sensors include strain sensors , proximity sensors , as well as image cameras and video cameras .

[0089] In another embodiment of the first aspect , after a period of time , preferably after a period of time from 1 min to 60 days , more preferably from 10 min to 30 days , even more preferably from 30 min to 15 days , most preferably from 60 min to 1 day the method as described herein is reiterated . This makes it possible to monitor the operating conditions over a longer period of time , for example to recognise trends in the data features . Typically, the period of time is in the order of 1 min, 5 min, 10 min, 30 min, 60 min, 2 h, 6 h, 12 h, 24 h, 2 d, 5 d, 10 d, 30 d, or 60 d .

[0090] In a preferred embodiment of the first aspect , the dataset comprises a time derivative of the time-series sensory data as a function of the time-series sensory data, according to Formula 1 : ( Formula 1 ) , wherein dy / dt is the time derivative of the time-series sensory data and y is the time-series sensory data . This functional relationship can be represented graphically by plotting the timeseries sensory data on the X-axis and the time derivative thereof on the Y-axis of a plot .

[0091] The inventors have found that the functional relationship according to Formula 1 and data features extracted from it can be used to improve the condition monitoring of reciprocating compressors . In particular, an improvement was observed in comparison to methods in which data features are extracted from the time derivative of time-series sensory data as a function of time according to Formula 2 :

[0092] — = f(t') ( Formula 2 ) , dtJ k Jwherein t is the time .

[0093] In another preferred embodiment of the first aspect , the dataset comprises a time derivative of the time-series sensory data as a function of the time-series sensory data, and the dataset comprises a mathematical set 1:

[0094] (set 1) , wherein y(ti) corresponds to a data point of the time-series sensory data (1) at a time point t±, and dy / dt(ti) corresponds to a data point of the time derivative of the time-series sensory data at the same time point t± .

[0095] By way of a simplified example, set 1 may include the following datapoints :

[0096] {(100,2), (150,5), (200,8), (250,12), (300,15), (310,5), (300, -2), (250,-10), (200, -8)} .

[0097] Each pair of datapoints comprises a first value for the pressure in bar and a second value for the time derivative of the pressure, both at the same time point. Data features are extracted from this set in step c. of the method according to the first aspect. The inventors have found that extracting data features in step c. from such a dataset enables improved condition monitoring of reciprocating compressors.

[0098] A second aspect of the present invention relates to a computer- implemented method for determining at least one cause of an emerging malfunction or a malfunction of a reciprocating compressor. The method comprises the following steps: In step a., an operating condition of the reciprocating compressor is determined according to the method of the first aspect of the invention. The operating condition is selected from the group consisting of a normal operating condition, an operating condition with an emerging malfunction, and an operating condition with a malfunction. In step b . , if the operating condition is not a normal operating condition, a cause analysis on the set of extracted data features is performed to determine the at least one cause of the emerging malfunction or the malfunction. Exemplarily, the cause analysis may comprise an analysis of the deviations in multiple extracted data features . One method is to utili ze pattern recognition, classi fication, or clustering algorithms to identi fy the underlying issues . By examining how these features deviate together, it is possible to link speci fic deviation patterns to particular component failures or operational problems . Firstly, deviations in multiple extracted data features originating from one or more time-series sensory data types such as pressure , temperature , vibration, and / or acoustic signals are analysed . These deviations are then aggregated into a multi-dimensional space where each dimension represents a di fferent feature . Advanced pattern recognition techniques , such as machine learning classi fiers ( e . g . , support vector machines , decision trees ) or clustering algorithms ( e . g . , k-means clustering, hierarchical clustering) , can be applied to this multi-dimensional data . The algorithms are trained on historical data where the causes of non-normal conditions were previously identi fied, allowing them to learn the characteristic patterns associated with di f ferent types of failures . Once trained, these algorithms can classi fy or cluster new sets of deviations to diagnose the speci fic cause of a non-normal operating condition . For example , i f a certain pattern of a change in the maximum compression rate and the y-intercept of the linear approximation of the compression phase is consistently associated with bearing wear, the algorithm can recognise this pattern in the current data and suggest bearing wear as the probable cause .

[0099] A third aspect of the invention relates to a data processing system comprising means for carrying out the method of the first and / or second aspect of the invention .

[0100] A fourth aspect of the invention relates to a reciprocating compressor comprising at least one sensor type (preferably selected from the group consisting of pressure sensors , temperature sensors , vibrational sensors , acoustic sensors , strain sensors , proximity sensors , and cameras ) to acquire time-series sensory data (preferably selected from the group consisting of pressure data, temperature data, vibrational data, acoustic data, strain data, position data, proximity sensor data, image data, and video data ) . The reciprocating compressor further comprises the data processing system of the third aspect of the invention; or a data connection to the data processing system of the third aspect of the invention to trans fer at least a part of the acquired time-series sensory data . The data connection is wireless or wired .

[0101] A fi fth aspect of the invention relates to a computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of the first and / or second aspect of the invention .

[0102] A sixth aspect of the invention relates to a computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out the method of the first and / or second aspect of the invention .

[0103] The invention will now be described by way of more speci fic embodiments . These embodiments are not intended to limit the gist of the invention in any way, but rather serve to ease understanding of the invention .

[0104] Fig . 1 : simpli fied flow diagram of the method according to the invention; Fig. 2: time-series sensory data of the cylinder pressure of a compressor stage during several compression cycles of a reciprocating compressor;

[0105] Fig. 3: first time derivative of the time-series sensory data from Fig. 2;

[0106] Fig. 4: visual representation of a dataset created in step b. of the method according to the invention comprising a time derivative of the time-series sensory data as a function of the time-series sensory data;

[0107] Fig. 5: visual representation of the dataset from Fig. 4 annotated with data features extracted from the dataset; and

[0108] Fig. 6A-D: visual representation of an evolution of the dataset from Fig. 4 over time, initially showing a normal operating condition (Fig. 6A and 6B) , then evolving into an operating condition with an emerging malfunction (Fig. 6C) and finally ending in an operating condition with a malfunction (Fig. 6D) .

[0109] The present example demonstrates the determination of an operating condition of a single-acting reciprocating compressor using a measurement of the cylinder pressure over time. To this end, the cyclical variations in pressure in the corresponding compressor cylinder are first recorded at a specific point in time using a pressure sensor as it is known in the art. The number of time-series data points recorded is such that the data from slightly more than three crankshaft revolutions, i.e. compression cycles, are recorded. The measurement data obtained is transmitted to a data processing system as time-series sensory data 1 . It is irrelevant for the present invention at which points within a compression cycle the data series begins and ends . The start and end of the compression cycles do not have to be determined and / or marked either in the sensor or during the subsequent data evaluation . It is suf ficient to use the current compressor speed to determine the required duration in which data points are recorded in order to record slightly more than three compression cycles , for example .

[0110] Figure 1 shows a simpli fied flow diagram of the method according to the invention . The time-series sensory data 1 are received 100 by the data processing system . The time-series sensory data 1 are represented in Figure 2 and the first time derivative 2 of those data in Figure 3 . A dataset 3 is then created 101 comprising the time derivative of the time-series sensory data 2 as a function of the time-series sensory data 1 . A visual representation of this dataset 3 is plotted in Figure 4 . Such a plot is not necessarily created when the subsequent analysis is solely a computational analysis without user interaction, which would also be possible . However, the present example includes both a computational analysis and a user analysis , as described in more detail below . The slightly more than three compression cycles from Figures 2 and 3 are shown in Figure 4 , one above the other, in the form of an orbit plot . The temporal progression of the recorded data runs clockwise in this representation . This illustration also makes it clear why it is not necessary to determine the exact start and end points of a compression cycle , as such an orbit plot does not include a start or end . As described above , this simpli fies the commissioning of condition monitoring, as di f ferent measured values do not have to be synchronised via a trigger signal . The two data series ( time-series data 1 and its time derivative 2 ) are always inherently in sync, even i f the compressor speed changes . The dataset 3 , which is used for the subsequent analysis , is also not dependent on other data such as the change in cylinder volume or the crank angle , which are time-consuming to determine and prone to errors . In the present example , no time-delay embeddings are used .

[0111] A set of data features is extracted 102 from the dataset 3 . The data points of the slightly more than three compression cycles are used collectively to determine the data features . On the visual representation of the dataset 3 shown in Figure 4 (without data features ) and Figure 5 (with annotated data features ) two data ranges can be identi fied, which represent an approximately linear relationship between cylinder pressure ( 1 ) and the time derivative of the cylinder pressure ( 2 ) . The first range corresponds to the compression phase of the compression cycle and its linear approximation 6 has a positive slope 4 . The y-in- tercept 5 is approximately -2 bar / s . The second range corresponds to the values recorded during the expansion phase of the compression cycle . The linear approximation 9 of this has a negative slope 7 and the y-intercept is approximately -3 bar / s . These four values form the first data features that can be extracted from dataset 3 . In comparison, a p-V diagram of the prior art could also be used to determine a data range for the compression phase and a data range for the expansion phase . In such a p-V diagram, however, no linear dependence consists between the pressure and the cylinder volume , which would make the reproducible extraction of data features very di f ficult .

[0112] Further, the maximum compression rate , i . e . the maximum time derivative data point 10 of the time derivative , is extracted as a data feature . Finally, the area within the orbit plot is determined . In this example , the suction pressure ( Ps) 13 and the discharge pressure ( PD) 14 of the corresponding compressor stage are used as the limits of integration, which are approximately 98 bar and 298 bar respectively . To arrive at the area within the orbit plot ( grey area 11 ) and within these limits of integration, a first integral of the compression phase data within these limits is calculated, from which a second integral of the expansion phase data within these limits is subtracted . This represents the data feature of the di f ference 11 between a first integral value and a second integral value .

[0113] After extracting 102 the data features , an analysis of them is performed 103 . In this example , this analysis includes both a computational analysis and a user analysis . The method according to the invention is carried out for the first time at a first point in time . The data 1 and 2 from dataset 3 is shown in Figure 6A. The data features are extracted as described above 102 . To perform 103 the analysis , the orbit plot in Figure 6A is displayed to a user via a user interface together with annotated data features and the values determined therefrom . The corresponding compressor has been manufactured recently and the device test has shown that the operating condition of the compressor is normal . The user is requested to inform the system via a user input that the measured values correspond to a normal operating condition .

[0114] The method will be repeated at a later date . Additional data can of course be recorded and analysed between the data shown in Figures 6A to 6D . Data 1 and 2 are shown in Figure 6B . The user is now shown both this newly measured and determined data 1 and 2 ( dashed line ) and the original data of the compressor from Figure 6A. This historical data serves as reference data 12 ( solid line ) . In addition, the user is again shown the data features as annotations ( see Figure 5 ) and the resulting numerical values . The corresponding values of the historical data are also displayed as reference values . This gives the user the opportunity to detect deviations between the current values and the reference values . Based on the data in Figure 6B, the user can see that although there are slight deviations between the measured values and the reference values , these deviations are still within an acceptable range . The user informs the system via a user input that the operating condition is normal . In addition to the user analysis , an automated analysis is also carried out , wherein the data processing device performs a programmed statistical analysis on the data features in order to determine an operating condition itsel f . The automated operating condition and the user operating condition are then compared with each other and a possible deviation between the two analyses is communicated to the user . In the present case , both analyses determine a normal operating condition .

[0115] At an even later point in time , the measured values in Figure 6C show a greater deviation from the originally measured values 12 . For example , the two slopes 4 and 7 , as well as the y-intercepts 5 and 8 of the two linear approximations 6 and 9 have changed . The maximum compression rate 10 has also decreased and the area 11 within the orbit plot is also smaller ( see Figure 5 ) . In this case , the user analysis shows that this is an operating condition with an emerging mal function . Although the compressor still achieves the required compression performance , it should be examined whether one or more components require maintenance . The automated analysis also shows that at least individual data features are outside the normal range and indicates an emerging mal function .

[0116] However, the emerging mal function warning was ignored and the compressor remained in operation without maintenance . A few weeks later, data 1 and 2 in dataset 3 changed noticeably ( Figure 6D) , so that the compressor is no longer functional . This is also made clear by the data features determined, which now indi- cate an operating condition with a mal function through both the user analysis and the computational analysis . Maintenance of the af fected components is now unavoidable .

[0117] The preceding description serves as an example of monitoring an operating condition of a reciprocating compressor that deteriorates over time using the method according to the invention .

Claims

Claims1. A computer-implemented method for determining an operating condition of a reciprocating compressor, the method comprising the steps of: a. receiving (100) time-series sensory data (1) of the reciprocating compressor; b. creating (101) a dataset (3) comprising at least one of :- a time derivative of the time-series sensory data (2) as a function of the time-series sensory data ( 1 ) ; and- time-delay embeddings of the time-series sensory data; c. extracting (102) a set of data features from the dataset ( 3 ) ; d. performing (103) an analysis on the set of extracted data features to determine and operating condition of the reciprocating compressor.

2. The method of claim 1, wherein the time-series sensory data (1) exhibit cyclic variations depending on a revolution of a crankshaft of the reciprocating compressor.

3. The method of any one of the preceding claims, wherein the time-series sensory data (1) are selected from the group consisting of pressure data, temperature data, vibrational data, acoustic data, strain data, position data, proximity sensor data, image data, and video data; preferably selected from the group consisting of pressure data, temperature data, and acoustic data.

4. The method of any one of the preceding claims, wherein the time-series sensory data (1) comprise or are pressure data, preferably pressure data selected from the group consisting of suction pressure data, discharge pressure data, cylinder pressure data, interstage pressure data, and crankcase pressure data, more preferably the pressure data are cylinder pressure data.

5. The method of any one of the preceding claims, wherein the time-series sensory data (1) do not comprise at least one sensory data type from the group consisting of displacement data, crank angle data and volume data.

6. The method of any one of the preceding claims, wherein the dataset (3) created in step b. comprises a time derivative of the time-series sensory data (2) as a function of the time-series sensory data (1) , and each data feature extracted in step c. is of any of the following data feature types :- the slope (4) and / or the y-intercept (5) of a first linear approximation (6) of a first set of data points of the dataset, preferably a linear approximation (6) of a set of data points of the dataset representing at least a part of the compression phase of a compression cycle ;- the slope (7) and / or the y-intercept (8) of a second linear approximation (9) of a second set of data points of the dataset, preferably a linear approximation of a set of data points of the dataset representing at least a part of the expansion phase of a compression cycle;- the maximum time derivative data point (10) and, optionally, the corresponding time-series data point;- a first integral value of a third set of data points of the dataset, preferably an integral value of a set of data points of the dataset representing at least a part of the compression phase of a compression cycle;- a second integral value of a fourth set of data points of the dataset, preferably an integral value of a set of data points of the dataset representing at least a part of the expansion phase of a compression cycle;- the difference (11) between the first integral value and the second integral value.

7. The method of any one of the preceding claims, wherein- the time-series sensory data (1) are pressure data, preferably cylinder pressure data;- the dataset (3) created in step b. comprises the pressure derivative (2) as a function of the pressure data (1) ; and- the data features extracted in step c. are at least one of, preferably at least the majority of, more preferably all but one of:- the slope (4) and / or the y-intercept (5) of a linear approximation (6) of the pressure derivative (2) as a function of the pressure data (1) representing at least a part of the compression phase of a compression cycle;- the slope (7) and / or the y-intercept (8) of a linear approximation (9) of the pressure derivative (2) as a function of the pressure data (1) representing at least a part of the expansion phase of a compression cycle ;- the maximum pressure derivative value (10) and, optionally, the corresponding pressure value;- a first integral value of the pressure derivative as a function of the pressure data representing at least a part of the compression phase of a compression cycle ;- a second integral value of the pressure derivative as a function of the pressure data representing at least a part of the expansion phase of a compression cycle ;- the di f ference ( 11 ) between the first integral value and the second integral value .8 . The method of any one of the preceding claims , wherein the analysis performed ( 103 ) in step d . comprises a user analysis comprising the steps of :- displaying the extracted data features and / or the received time-series sensory data ( 1 ) to a user on a user interface ;- optionally, displaying reference values ( 12 ) to the user on the user interface ;- receiving a user input .9 . The method of any one of the preceding claims , wherein the analysis performed ( 103 ) in step d . comprises a computational analysis , and- performing the computational analysis on the extracted data features comprises applying a trained machine-learning model on the extracted data features ; and- optionally, the machine-learning model is trained using an unsupervised, a semi-supervised, or a supervised learning algorithm, preferably at least one of- an ensemble learning algorithm, preferably at least one of- a boosting algorithm, preferably a gradient boosting algorithm; and- a random forest algorithm; and- a neural network algorithm .10 . The method of any one of the preceding claims , wherein the analysis performed ( 103 ) in step d . comprises a computational analysis , and performing the computational analysis on the extracted data features comprises applying a weighted average formula on the extracted data features .11 . The method of any one of the preceding claims , wherein the operating condition is selected from the group consisting of :- a normal operating condition;- an operating condition with an emerging mal function; and- an operating condition with a mal function .12 . The method of any one of the preceding claims , comprising at least one additional step selected from the group consisting of :- noti fying a user of the determined operating condition and / or ( a ) determined deviation ( s ) ;- prompting a user to take measures to restore a normal operating condition and / or to stop at least part of the reciprocating compressor ; and- sending instructions to a control system of the reciprocating compressor to stop at least part of the reciprocating compressor ; i f the determined operating condition does not correspond to the normal operating condition .

13. The method of any one of the preceding claims, comprising, before step a., an additional step of acquiring time-series sensory data (1) using at least one sensor, preferably selected from the group consisting of pressure sensors, temperature sensors, vibrational sensors, and acoustic sensors .

14. The method of any one of the preceding claims, wherein after a period of time, preferably after a period of time from 1 min to 60 days, more preferably from 10 min to 30 days, even more preferably from 30 min to 15 days, most preferably from 60 min to 1 day the method of any one of the preceding claims is reiterated.

15. The method of any one of the preceding claims, wherein the dataset (3) comprises a time derivative of the time-series sensory data (2) as a function of the time-series sensory data (1) , according to Formula 1:(Formula 1) , wherein dy / dt is the time derivative of the time-series sensory data (2) and y is the time-series sensory data (1) .

16. The method of any one of the preceding claims, wherein the dataset (3) comprises a time derivative of the time-series sensory data (2) as a function of the time-series sensory data (1) , and the dataset (3) comprises a mathematical set 1 :(set 1) ,wherein y(t±) corresponds to a data point of the time-series sensory data (1) at a time point t±, and dy / dt (ti) corresponds to a data point of the time derivative of the timeseries sensory data (2) at the same time point t± .17 . A computer-implemented method for determining at least one cause of an emerging mal function or a mal function of a reciprocating compressor, the method comprising the steps of : a . determining an operating condition of the reciprocating compressor according to the method of any one of claims 1 to 16 , wherein the operating condition is selected from the group consisting of a normal operating condition, an operating condition with an emerging mal function, and an operating condition with a malfunction; b . i f the operating condition is not a normal operating condition, performing a cause analysis on the set of extracted data features to determine the at least one cause of the emerging mal function or the mal function .18 . A data processing system comprising means for carrying out the method of any one of claims 1 to 17 .19 . A reciprocating compressor comprising at least one sensor type , preferably selected from the group consisting of pressure sensors , temperature sensors , vibrational sensors , acoustic sensors , strain sensors , proximity sensors , and cameras , to acquire time-series sensory data ( 1 ) , preferably selected from the group consisting of pressure data, temperature data, vibrational data, acoustic data, strain data, position data, proximity sensor data, image data, and video data ; and- the data processing system of claim 18 ; or- a data connection to the data processing system of claim 18 to trans fer at least a part of the acquired time-series sensory data .20 . A computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of any one of claims 1 to 7 , 9 to 12 , or 14 to 17 .21 . A computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out the method of any one of claims 1 to 7 , 9 to 12 , or 14

Images