Root cause analysis of anomalies in turbomachines

Abstract

The innovative method allows the determination of a root cause of an anomaly in a turbomachine and comprises the steps of: a) receiving (220) measurement data relating to a time frame from a set of feature sensors mounted on the turbomachine, b) receiving (230) an identifier of a target feature sensor in which an anomaly occurs, and c) determining (230) the identifier of the target feature sensor. C) receiving (240) a start time and an end time of a non-exception time subframe, d) receiving (250) a start time and an end time of an exception time subframe, e) receiving (260) identifiers of a plurality of feature sensors associated with features of the turbine that may be a root cause of the exception, and f) deriving (270) at least one feature of the turbine considered to be a root cause of an anomaly based on a comparative analysis of the feature importance values of the feature sensor during the non-abnormal time subframe and during the abnormal time subframe.

Description

[0001] Specification Technical Field

[0002] The subject matter disclosed herein relates to root cause analysis of anomalies in turbines. Background Art

[0003] Even though turbines are designed to avoid anomalies and maintenance is intended to prevent anomalies, anomalies do occur in turbines during operation. An anomaly is a condition in a machine that is away from standard operability; a first example of an anomaly is vibration at a specific location in the machine, which has an amplitude higher than the normal vibration amplitude at that location; a second example of an anomaly is the rotational speed of a certain component of the machine being higher than the normal rotational speed amplitude of that component; a third example of an anomaly is the temperature at a certain location in the machine having a value higher than the normal temperature at that location; a fourth example of an anomaly is the pressure in a specific pipe or chamber of the machine having a value higher than the normal pressure in that pipe or chamber; a fifth example of an anomaly is the flow rate in a specific pipe of the machine having a value higher than the normal flow rate in that pipe or chamber. The expression "away from" should be interpreted to mean that the difference between the standard value (e.g., the rated value) and the actual value is greater than a pre-determined difference, e.g., a pre-determined percentage difference; such pre-determined differences typically vary depending on the parameter, and can also depend on, for example, the operating state of the machine.

[0004] Root cause analysis of anomalies in turbines, i.e., finding the reasons why anomalies have occurred in the past or are currently occurring in turbines, is very important for both manufacturers and users, but is very difficult to perform reliably. The complexity of turbines (such as compressors or turbines) and the complexity of the competition in, for example, oil and gas applications, in which the turbines are installed and operated, make this task even more difficult. In some cases, it is even difficult to precisely identify the anomaly.

[0005] The technical literature discloses computer-implemented methods and computer-based systems aimed at fully automatically identifying anomalies in machines. Some anomalies are relatively easy to identify. Other anomalies are difficult to identify. For effective and reliable execution, a common possibility is to perform extensive testing and training on each machine of interest. Generally, in-depth knowledge of the machine of interest and of the competition in which the machine of interest is installed and operated is a great advantage for a reliable solution.

[0006] Similarly, the technical literature discloses computer-implemented methods and computer-based systems aimed at fully automatically finding the root causes of anomalies in machines. Such a task is much more difficult, and the effectiveness and reliability are much more challenging. Even in this case, extensive testing and training can be used to solve the problem.

[0007] Accordingly, there is a desire for an easier way to analyze the root cause of anomalies in turbines, specifically turbines for oil and gas applications, without sacrificing effectiveness and reliability. SUMMARY OF THE INVENTION

[0008] According to a first aspect, the subject matter disclosed herein relates to a computer-implemented method for root cause analysis of anomalies in a turbine; the method comprising the steps of: a) receiving measurement data associated with a time frame from a set of feature sensors mounted on the turbine, b) receiving an identifier of a target feature sensor in whose measurement data the anomaly occurs, c) when the anomaly does not occur, receiving the start time and end time of a non-anomaly time sub-frame, d) when the anomaly does occur, receiving the start time and end time of an anomaly time sub-frame, e) receiving identifiers of a plurality of feature sensors associated with features that may be the root cause of the anomaly in the turbine, and f) deriving at least one feature of the turbine that is considered to be the root cause of the anomaly based on a "contrast analysis" of "feature importance" values of the feature sensors during the non-anomaly time sub-frame and the anomaly time sub-frame. The terms "contrast analysis" and "feature importance" will be explained in the detailed description below.

[0009] According to other aspects, the subject matter disclosed herein relates to computer-based systems and turbine arrangements in which such methods are implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] When considered in conjunction with the accompanying drawings, the disclosed embodiments of the invention and many of its attendant advantages become better understood, and thus will be readily appreciated more fully, wherein:

[0011] Figure 1 A schematic block diagram showing one embodiment of an innovative turbine arrangement including an innovative system,

[0012] Figure 2 A flowchart showing an embodiment of an innovative method for root cause analysis of anomalies in a turbine, and

[0013] Figure 3 Showing Figure 2 A flowchart of a possible embodiment of a particular step of the method. DETAILED DESCRIPTION

[0014] As described above, it is difficult to identify anomalies in a turbine if reliable results are required, and it is even more difficult to determine their root causes. Therefore, it has been envisaged to limit the task to a simplified, still challenging problem, but to avoid preliminary testing and training of one or more machines. It is assumed that A) anomalies are identified, i.e., the anomalous time periods and the non-anomalous time periods are known, and B) the possible causes of the anomalies are known. The task is to select the best cause, i.e., the cause that is likely the true root cause of the anomaly. Typically, the number of possible causes of an anomaly is large, e.g., from e.g. 10 to e.g. 100, and depends on the different anomalies; for example, a vibration above normal may be caused by an anomalous flow value in any one of a set of pipes, an anomalous pressure value in any one of a set of pipes, an anomalous rotational speed of any one of a set of components; the task is to select which one of the pipe flow or the pipe pressure or the component rotational speed is causing or is causing the anomaly identified at a particular time. According to the subject matter disclosed herein, the problem is solved by "contrast scores" (i.e., comparing the anomalous time periods and the non-anomalous time periods), and does not require any preliminary knowledge of any anomalies, specifically it does not require preliminary training.

[0015] Reference will now be made in detail to embodiments of the present disclosure, examples of which are illustrated in the accompanying drawings. The examples and the drawings are provided in a manner to explain the present disclosure and should not be construed as limiting the present disclosure. In fact, it will be apparent to those skilled in the art that various modifications and variations can be made to the present disclosure without departing from the scope or substance thereof. In the following description, like reference numerals are used for the drawings of the embodiments to indicate elements performing the same or similar functions. Additionally, for clarity of illustration, some references may not be repeated in all the drawings.

[0016] In Figure 1 a very schematic manner, an embodiment of the innovative turbine arrangement 100 and its user 10 interacting with the arrangement 100 are shown. The arrangement 100 includes a turbine 180 and an innovative computer-based system 140. The user 10 can be an employee of a company manufacturing the turbine 180, or an employee of a company responsible for testing the turbine 180, or an employee of a company managing the plant in which the turbine 180 is installed. More generally, the user 10 is a person or a team of persons interested in determining the root cause of an anomaly occurring or taking place in the turbine 180; such a need may be repeated from time to time at any new anomaly.

[0017] Arrangement 100 and its variants will be described in more detail hereinafter. For now, it is important that turbine 180 is expected to include a set of characteristic sensors 182 mounted thereon; the number of sensors 182 is large, e.g., 100 - 1000; the sensors 182 repeatedly perform measurements of "characteristics" of the turbine, which "characteristics" may also be referred to as "variables", such as, for example, temperature, pressure, volume and mass flow rate, displacement, velocity (e.g., rotational speed), acceleration, vibration, valve opening, IGV set angular position, IGV detected angular position, gas composition, burner state. Generally, sensors 182 are "real" sensors, i.e., devices that perform measurements inside the turbine and determine / output (analog or digital) signals whose amplitude corresponds to the measured value. Alternatively, according to the subject matter disclosed herein, one or more of the sensors may be so-called "virtual" sensors; as is well known, a "virtual" sensor is software running in a computer (which may be the same computer that executes the method of the present invention), which uses input data from one or more "real" sensors and generates output data of the "virtual" sensor to repeatedly calculate, for example, formulas as if the machine would have an on-board "real" sensor instead of a "virtual" sensor.

[0018] System 140 receives measurement data from sensors 182 in some way. Figure 1 An arrow between turbine 180 and system 140 is shown, which can be interpreted as a (wired or wireless) connection such that measurement data is received directly from turbine 180. However, according to some embodiments, the measurement data may be collected, for example, by a computer system ( Figure 1 (not shown in the figure) at some time and then transferred to computer system 140 at a later time via a (wired or wireless) computer connection or via a data storage device, solely for analysis.

[0019] Figure 2 A flowchart 200 of an embodiment of an innovative method for root cause analysis of anomalies in a turbine (e.g., Figure 1 turbine 180 in the figure) is shown; it is a computer-implemented method that can be implemented, for example, by Figure 1 computer system 140 in the figure. At each identified anomaly, the path of the flowchart goes from start box 210 to stop box 280 -- generally speaking, it is expected that during the operation of the turbine, several anomalies usually occur one after another; thus, it can be repeated, for example, during the study or inspection of the turbine (i.e., offline) or during the test of the turbine or during the operation of the turbine.

[0020] The innovative method includes the following steps:

[0021] a) Receive (block 220) data from a set of feature sensors installed on a turbine, the data corresponding to measurements performed by the feature sensors in the set of feature sensors within a time frame,

[0022] b) Receive (block 230) an identifier of a feature sensor in the set of feature sensors, where the anomaly appears in measurement data from at least the feature sensor and the feature sensor is the target feature sensor of the anomaly,

[0023] c) Receive (block 240) a first start time and a first end time of a first time sub-frame, the first time sub-frame being included in the above-mentioned time frame, where the anomaly does not occur during the first time sub-frame,

[0024] d) Receive (block 250) a second start time and a second end time of a second time sub-frame, the second time sub-frame being included in the above-mentioned time frame, where the anomaly does occur during the second time sub-frame,

[0025] e) Receive (block 260) identifiers of a plurality of feature sensors in the set of feature sensors associated with features of the turbine that may be the root cause of the anomaly, and

[0026] f) Derive (block 270) at least one feature of the turbine that is considered to be the root cause of the anomaly based on a "contrast analysis" of the "feature importance" values of the feature sensors in the plurality of feature sensors during the first time sub-frame and the second time sub-frame.

[0027] The "contrast analysis" in step "f" means comparing the anomaly time period with the non-anomaly time period, specifically the "feature importance" at the anomaly time period and the non-anomaly time period; step "f" will be assisted by

[0028] Figure 3 ​It will be better explained later. The term "feature importance" means the degree or level of the influence of an input feature (i.e., a variable) on an output feature (i.e., a variable). If the system is considered as a black box with several input variables and one output variable, any specific output value can be considered due to the influence of all input values; however, the contribution of each input value to a specific output value may be different. This system can be associated with a usually very complex "prediction model" and an "interpretation model", which should be very simple for easy understanding. A very effective type of "interpretation model" is a linear function of binary variables, such that an "additive feature attribution method" can be implemented. More details on this topic can be found, for example, in the article named "A unified approach to interpreting model predictions" by Scott M. Lundberg and Su-In Lee in the proceedings of the 31st International Conference on Neural Information Processing Systems - NIPS 2017.

[0029] Specifically, in step "f", at least one feature is selected within a plurality of features (i.e., the features associated with the plurality of feature sensors mentioned at step "e"). Specifically, at step "f", the feature importance difference of each feature sensor among the plurality of feature sensors is calculated, and the highest feature importance difference is determined therefrom. The reason why more than one feature can be obtained at step "f" will be explained later and is related to the fact that this innovative method can be used as an aid for a human (or a human team, such as a technical expert), such that the final decision on which one of them is the real root cause may be left, for example, to such a human (or a human team, such as a technical expert), who also based on their expertise or may need further testing and / or investigation.

[0030] The "contrast analysis" at step "f" according to the subject matter disclosed herein includes:

[0031] - During the first time sub-frame and the second time sub-frame, calculating the feature importance value of each feature sensor among the plurality of feature sensors, specifically, according to the feature contribution to the target variable,

[0032] - Calculating the feature importance difference of each feature sensor among the plurality of feature sensors, where the feature importance difference is the difference between the feature importance value during the first time sub-frame and the feature importance value during the second time sub-frame, and

[0033] - Determining the high feature importance difference or the highest feature importance difference among the calculated feature importance differences.

[0034] Advantageously, the feature importance values are calculated based on a turbine model (usually an interpretive model) that is a linear function of binary variables corresponding to features of the turbine, where the features of the turbine correspond to the plurality of feature sensors, as mentioned in the already cited article by Scott M. Lundberg and Su-In Lee, for example.

[0035] According to some exemplary embodiments, one or more or all of the "receiving" steps (i.e., steps "b" and "c" and "d" and "e") include receiving an input from a user (e.g., Figure 1 user 10 in

[0036] According to some exemplary embodiments, the occurrence of an anomaly is judged by a human based on human observations of the turbine and, for example, its measurement data.

[0037] Typically, a second time subframe (i.e., the "anomaly" time subframe) follows a first time subframe (i.e., the "non-anomaly" time subframe). In fact, the anomaly may have even started at the end of the first time subframe, but its effects are not obvious during the first time subframe, not even at the end of the first time subframe. If the human makes an incorrect judgment, the innovative method still provides good results if only a few measurement data in the first time subframe are collected when the anomaly occurs.

[0038] In the above steps "b" and "e", "identifiers" are referred to as means for identifying the sensors and features of the turbine.

[0039] In the above steps "c" and "d", "start time" and "end time" are referred to as means for identifying the "time subframe". Equivalently, the "time subframe" can be identified, for example, by "start time" and "duration" or "end time" and "duration".

[0040] According to an advantageous embodiment, for example Figure 3 the embodiment of

[0041] f1) Based on the received measurement data (specifically, the measurement data in the first time subframe (i.e., the "non-anomaly"

[0042] time subframe)) create (block 272) a model (usually an interpretive model) of the turbine that has at least the features of the turbine corresponding to the plurality of feature sensors as inputs and at least the features of the turbine corresponding to the target feature sensor as outputs,

[0043] f2) Based on the model created at sub-step "f1", during this first time sub-frame and this second time sub-frame, calculate (block 274) the feature importance value of each of the plurality of feature sensors relative to the measurement data from at least the target feature sensor.

[0044] f3) Based on the feature importance values calculated at sub-step "f2", calculate (block 275) the feature importance difference of each of the plurality of feature sensors, where the feature importance difference is the difference between the feature importance value during the first time sub-frame and the feature importance value during the second time sub-frame.

[0045] f4) Determine (block 276) the highest feature importance difference among the feature importance differences calculated at sub-step "f3", and

[0046] f5) Derive (block 278) from the highest feature importance difference determined at sub-step "f4"

[0047] the associated feature of the turbine that is considered to be the root cause of the anomaly.

[0048] The model at sub-step "f1" is advantageously a regression model, which more advantageously can be implemented by a recurrent neural network, specifically a neural network of the LSTM (= "Long Short-Term Memory") type. Such a model (specifically a neural network) is trained to predict the target feature based on the measurement data of the input feature sensors using the data received at step "a".

[0049] According to an advantageous embodiment, the sub-step "f2" is implemented by an interpretability method that is applied on top of the model created at sub-step "f1" and is based on Shapley value-related techniques, specifically SHAP values (see, for example, the already cited article by Scott M. Lundberg and Su-In Lee, which provides a general explanation of Shapley value-related techniques and a specific description of SHAP values). Such interpretability methods originate from game theory and essentially assign a value to each feature that corresponds to the change in the expected model prediction when that feature is adjusted. The basic procedure behind such interpretability methods is to retrain the model on all possible feature subsets S of F, where F is the set of all features, and assign an importance value to each feature that represents the impact of including that feature on the model prediction. To calculate this effect, the model is trained in the presence of the feature and another model is trained while withholding the feature. Since the effect of suppressing a feature depends on the other features in the model (collinearity effects), the aforementioned difference is calculated for all possible feature subsets. Then the Shapley values are calculated and used as feature attributes. They are the weighted average of all possible differences in the predictions obtained with and without a specific feature. Usually, this calculation is approximated in order to speed up the process.

[0050] Preferably, the operating conditions of the turbine in the first time sub-frame and the operating conditions of the turbine in the second time sub-frame are similar. The similarity can be based on the values of the input features; for example, similar conditions can mean, for the compressor, within the same rotational speed range and / or within the same suction pressure range and / or within the same discharge pressure range. The similarity can be based on the values of the output features, i.e., based on the effects; the same conditions can mean that the target feature should have the same value or be within the same value range if no anomalies should be present. From a practical point of view, the similarity can be based on temporal proximity; if the first time sub-frame and the second time sub-frame are consecutive or close to each other in time (e.g., the time distance is less than 10% or 20% or 50% or 100% of the duration of the first time sub-frame or the second time sub-frame), then similarity is possible.

[0051] According to some advantageous embodiments:

[0052] At sub-step "f4", a set of the highest feature importance differences is determined, and

[0053] At sub-step "f5", the feature importance differences in the set of the highest feature importance differences are sorted, and the corresponding associated features of the turbine are sorted according to the root cause of the anomaly.

[0054] In this way, some features are provided as possible root causes of the anomaly, and they are ranked according to the likelihood of being the true root cause.

[0055] According to some advantageous embodiments, at sub-step "f5", a confidence value is determined for determining the root cause of the anomaly. Specifically, sub-steps "f1", "f2", "f3", "f4", and "f5" are repeated based on different initializations of the weights of the neural network; for each repetition, a different feature ranking is obtained; for each feature, the average of its ranking positions and the variance of its ranking positions are determined; the feature confidence as the root cause of the analysis is the reciprocal of its variance.

[0056] The method for root cause analysis of anomalies in turbines described and claimed herein can be implemented by a computer-based system, such as Figure 1 the system 140 in Figure 1 is configured to execute the method. The system may generally include a processor (such as Figure 1 the processor 142 in Figure 1 ), a memory connected to the processor 142 and configured to store programs and data (such as

[0057] the memory 146 in

[0058] ), and a human I / O interface connected to the processor 142 (such as the human I / O interface 144). These components 142, 144, and 146 are key components of a computer; thus, the system 140 can be, for example, a so-called "workstation" or a so-called "server" or even a so-called computer "cluster". To execute the innovative method, an appropriate computer program is stored in the memory. To execute the method, an input from the user 10 is received from the human I / O interface and sent to the processor. As already described, the innovative system is configured to receive measurement data from the turbine in a certain way (see, for example Figure 1 the arrow in Figure 1 ). A typical possibility is that the system 140 includes a database 148 for storing data (specifically, measurement data) from one or more turbines. It is known in the art to transfer measurement data from turbines to a computer located remotely from the turbines and store them in a database located inside or coupled to the computer, and this is outside the scope of protection of this patent application.

[0059] According to possible embodiments, an innovation system (such as Figure 1 system 140 in Figure 1 ) can be configured to automatically identify one or more (not necessarily all) anomalies or one or more (not necessarily all) types of anomalies in a turbine (such as Figure 1 turbine 180 in

[0060] As Figure 1 shown, the innovation system can be integrated in a turbine arrangement, such as Figure 1 arrangement 100 in Figure 1 ). Such a system basically includes a turbine, such as Figure 1 turbine 180 in

[0061] It should be understood that the root cause of an anomaly in a turbine is useful not only for obtaining more / better knowledge of that turbine and / or for better designing that turbine or similar turbines (specifically one or more turbine components). In fact, according to some embodiments, based on the identified root cause(s), when the turbine is operating (on-site or during testing), an alert / warning (auditory or visual) can be triggered for the user and / or some action can be taken, such as controlling the turbine or one or more subsystems of the turbine or a system connected to the turbine via a computer. The action can be motivated, for example, by safety concerns or efficiency goals. The action can be, for example, shutting down the turbine or activating a safety system or subsystem or changing the regulation of a component (e.g., opening or closing a valve).

Claims

1. A computer-implemented method for root cause analysis of anomalies in a turbine, wherein the method comprises the following steps: a) Receiving (220) data from a set of feature sensors installed on the turbine, the data corresponding to measurement results performed by the feature sensors in the set of feature sensors within a time frame; b) Receiving (230) identifiers of the feature sensors in the set of feature sensors, wherein the anomaly appears in the measurement data from at least the feature sensors, and the feature sensors are the target feature sensors of the anomaly; c) Receiving (240) a first start time and a first end time of a first time sub-frame, the first time sub-frame being included in the time frame, wherein the anomaly does not occur during the first time sub-frame; d) Receiving (250) a second start time and a second end time of a second time sub-frame, the second time sub-frame being included in the time frame, wherein the anomaly does occur during the second time sub-frame; e) Receiving (260) identifiers of a plurality of feature sensors in the set of feature sensors associated with features of the turbine that may be the root cause of the anomaly; and f) Deriving (270) at least one feature of the turbine that is considered to be the root cause of the anomaly based on a comparative analysis of the feature importance values of the feature sensors in the plurality of feature sensors during the first time sub-frame and the second time sub-frame; wherein the comparative analysis in step "f" (270) comprises: - Calculating, during the first time sub-frame and the second time sub-frame, the feature importance value of each feature sensor in the plurality of feature sensors; - Calculating the feature importance difference of each feature sensor in the plurality of feature sensors, the feature importance difference being the difference between the feature importance value during the first time sub-frame and the feature importance value during the second time sub-frame; and - Determining a high feature importance difference or the highest feature importance difference among the calculated feature importance differences.

2. The method according to claim 1, wherein the feature importance value is calculated based on a model of the turbine, the model being a linear function of binary variables corresponding to features of the turbine, and the features of the turbine corresponding to the plurality of feature sensors.

3. The method according to claim 1, wherein step f comprises the following sub-steps: f1) Creating (272) a model of the turbine based on the measurement data, the model having at least the features of the turbine corresponding to the plurality of feature sensors as inputs and at least the features of the turbine corresponding to the target feature sensors as outputs; f2) Calculating (274), based on the model created in sub-step "f1", during the first time sub-frame and the second time sub-frame, the feature importance value of each feature sensor in the plurality of feature sensors with respect to the measurement data from at least the target feature sensor. f3) Calculate (275) the feature importance difference of each of the plurality of feature sensors based on the feature importance values calculated in sub-step "f2", where the feature importance difference is the difference between the feature importance value during the first time sub-frame and the feature importance value during the second time sub-frame. f4) Determine (276) the highest feature importance difference among the feature importance differences calculated in sub-step "f3", and f5) Derive (278) the associated feature of the turbine that is considered to be the root cause of the abnormality from the highest feature importance difference determined in sub-step "f4".

4. The method according to claim 3, where the model in sub-step "f1" (272) is a regression model.

5. The method according to claim 4, where the model in sub-step "f1" (272) is implemented by a recurrent neural network.

6. The method according to claim 5, where the model in sub-step "f1" (272) is implemented by an LSTM type neural network.

7. The method according to claim 1, where sub-step "f2" (274) is performed by Shapley value related techniques.

8. The method according to claim 1, where the operating conditions of the turbine in the first time sub-frame and the operating conditions of the turbine in the second time sub-frame are similar.

9. The method according to claim 3, where in sub-step f4 (276), a set of the highest feature importance differences is determined, and where in sub-step f5 (278), the feature importance differences in the set of the highest feature importance differences are sorted, and the corresponding associated features of the turbine are sorted according to the root cause of the abnormality.

10. The method according to claim 3, where in sub-step f5 (278), a confidence value is determined for determining the root cause of the abnormality.

11. A computer-based system (140) configured to perform the root cause analysis method according to claim 1.

12. The computer-based system (140) according to claim 11, where the computer-based system is configured to perform the method according to claim 1 during the operation of the turbine (180).

13. The computer-based system (140) according to claim 12, where the computer-based system is configured to trigger an alarm / warning based on the root cause identified by the method.

14. The computer-based system (140) according to claim 12, where the computer-based system is configured to take actions on the turbine (180) or on one or more subsystems of the turbine (180) or on one or more subsystems coupled to the turbine (180) based on the root cause identified by the method.

15. A turbine arrangement (100), the turbine arrangement comprising a turbine (180) and a system (140) according to claim 11.

16. A turbine arrangement (100), the turbine arrangement comprising a turbine (180) and a system (140) according to claim 12.

Images