Hybrid risk model for maintenance optimization and system for performing such method

Abstract

A computer-implemented method for maintenance optimization of a fleet or group of turbomachinery assets is disclosed. The method comprises a model training and setup step aimed at setting configuration parameters that can be executed offline, and a step of online computation of new input data based on detected data and extracted statistical features. Subsequently, anomaly identification and classification are performed, calculating a risk assessment for estimating the risk of any event that the anomaly can cause requiring the execution of a maintenance task on one or more assets of the fleet.

Description

[0001] Description TECHNICAL FIELD

[0002] The present disclosure relates to a risk model implemented in a computer program, which is able to estimate the maintenance time using a digital computer-based analysis configured to identify anomalies on the signal / system / unit behavior, classify these anomalies and assign a severity gradient to the anomalies.

[0003] More specifically, the methods and systems for performing this new and useful analysis allow to manage, schedule and perform the maintenance of complex grouped turbomachinery assets, such as gas turbines, compressors and the like, in a new way. The commercial benefits of the solution provided herein favor the full optimization of the maintenance scenario (work scope and timeline) of one or more turbomachinery plants monitored by each user, and / or other constraints (maintenance time and cost, asset availability and reliability). BACKGROUND

[0004] The digital, computer-based scheduling and performance of the maintenance of complex grouped turbomachinery assets, such as gas turbines, compressors and their auxiliary systems, has become a necessary advantage for the users, owners and / or operators of such assets, whether installed in a plant or otherwise.

[0005] Modern technology-based allows to detect in almost real time the status of a running industrial and / or aeroderivative gas turbine by checking the operation of the sensors installed in the plant. In order to improve the maintenance of turbomachinery assets, it is desirable to predict any possible malfunction of the plant in order to increase the profitability of the service system, but in particular, in order to reduce any risk of downtime of the plant, which can also cause interruptions to the public service, with very high costs.

[0006] Several known data processing systems can be provided, which are connected or connectable to a gas turbine. Such systems are also able to connect with the digital operating system that controls such gas turbines and with the control system of the plant where such gas turbines are located, to receive or download the operating data for checking the status of the plant.

[0007] A processing center will typically process the downloaded data and, based on appropriate analysis, predict in advance any possible maintenance intervention in order to avoid any unscheduled downtime of the plant. A typical processing center includes servers and computers available on the edge or in the cloud, able to process the downloaded operating data of the gas turbine and / or of the plant by applying appropriate computer-based methods and algorithms, some of which can be provided by the original equipment manufacturer (OEM) and / or based on physical models and / or asset operability optimization models.

[0008] The standard approach for maintaining a scene modeling is to use a fixed time schedule method to plan maintenance activities.

[0009] However, using a single maintenance time estimation method (empirical, physics-based or data-driven) often fails to provide sufficient accuracy for event prediction and maintenance planning.

[0010] Due to reduced or missing prediction accuracy, the methods and algorithms on which the maintenance optimization methods available in the prior art are based negatively impact maintenance costs as well as asset reliability and availability.

[0011] Therefore, an improved optimization maintenance method and related processing system for making accurate predictions for maintaining turbomachinery equipment and related auxiliary systems would be truly appreciated in the art. SUMMARY

[0012] The availability of engineering knowledge, fleet event experience and monitoring data allows to propose a new and useful, digital, computer-based predictive maintenance service that can use or combine several methods to do so:

[0013] - an empirical method based on failure mode probability assessment;

[0014] - physics modeling for failure mode progression assessment; and / or

[0015] - a pure data-driven method for anomaly detection.

[0016] It is believed that only the Original Equipment Manufacturer (“OEM”) (as the owner of the monitoring data and asset design) has a unique advantage to combine all the above methods, ensuring increased accuracy when planning maintenance events for a complex turbomachinery asset fleet, such as a fleet of gas turbines and / or compressors, well in advance of any operational downtime risk. In order to optimize the maintenance scene, the accuracy of the estimation must be sufficient to provide the end user with effective insights by guaranteeing the expected reliability and availability. Combining the three methods allows to optimize the scope of the work by prioritizing activities based on asset knowledge and service data.

[0017] Therefore, the inventors propose a new and useful computer-based optimization method to manage, schedule and / or perform maintenance on complex turbomachinery asset fleets. This new method improves accuracy by combining OEM engine-specific physics modeling (OEM ownership) and fleet-based data-driven methods, which is challenging for non-OEM maintenance service providers as they have less operational and maintenance knowledge and less access to operational data from turbomachinery equipment.

[0018] In one aspect, the subject matter disclosed herein relates to a method for optimizing the maintenance of a complex fleet or a complex plant and / or related turbomachinery assets and their auxiliary systems. Each turbomachinery asset in the fleet is configured to generate one or more parameters acquired by sensors and / or computed by monitoring and control systems. The method comprises a model setup, wherein models are created and validated, and an online computation, wherein data acquired from the monitored fleet are processed and outputs are provided online. The online computation method comprises the steps of reading configuration parameters defined by the model setup, and acquiring / receiving operational signals outputted from sensors and / or from control systems and / or from monitoring and diagnostic services of turbomachinery assets over at least one time frame. The method further comprises the step of extracting statistical and / or mathematical features or parameters from the operational signals to detect one or more anomalies from the features or operational parameters of the signals according to the computation set by the configuration parameters. The classification of each anomaly allows to distinguish between system anomalies, so that a risk assessment step is performed for estimating the risk of any event leading to asset / system unserviceability and / or performance below contract agreement limits, and the need to perform maintenance tasks to restore asset / system serviceability and expected performance. A severity assignment step is performed on sensor anomalies to assign a severity to the anomalies identified as sensor faults. The maintenance actions are then assigned by a policy defined by the maintenance service owner and visualized in a monitor or web service, or sent as a file or through an alarm system, such as audio or video signaling devices, to users and / or field operators.

[0019] Further, another aspect of the subject matter disclosed herein is an offline model setup to configure the whole system and train the computation step.

[0020] In another aspect, the subject matter disclosed herein relates to a method for performing a risk model setup by determining asset health and abnormal behavior and related signal patterns and values based on OEM experience, historical fleet data and machine learning techniques.

[0021] In another aspect, the subject matter disclosed herein relates to the application of machine learning methods for anomaly detection, the method consisting of supervised and / or unsupervised techniques applied to identify abnormal features from reference health features.

[0022] In another aspect, the subject matter disclosed herein relates to the application of machine learning methods for anomaly detection, the method consisting of supervised and / or unsupervised techniques applied to identify abnormal features from reference health features.

[0023] In a further aspect, the subject matter disclosed herein relates to the application of machine learning methods for anomaly detection, the method consisting of supervised and / or unsupervised techniques applied to identify abnormal features from reference health features. BRIEF DESCRIPTION OF DRAWINGS

[0024] A more complete understanding of the disclosed embodiments and the many attendant advantages thereof will readily be had by reference to the following detailed description when considered in connection with the following drawings wherein:

[0025] Figure 1 is a block diagram of a system configured to optimize maintenance of a fleet of turbomachinery equipment;

[0026] Figure 2 is a block diagram of a central control unit according to an embodiment; and

[0027] Figure 3 is a flowchart of computer-implemented method steps to provide improved maintenance optimization for one or more turbomachinery equipment in a plant environment according to a first embodiment;

[0028] Figure 4 shows a set of representative curves of decorrelated signals from sensors installed on turbomachinery equipment according to Figure 1 ; and

[0029] Figure 5 shows a set of representative curves representative of processed signals from sensors installed on turbomachinery equipment according to Figure 1 , which deviate from regular operating behavior. DETAILED DESCRIPTION

[0030] The computer-readable and executable program estimates maintenance times for one or more turbomachinery assets and / or related equipment (drivers / drive equipment and / or auxiliary systems, as better explained below) in an industrial plant environment in conjunction and using models or algorithms. This model is the basis for an optimized maintenance service planning, aiming at minimizing asset downtime and maximizing plant production, as well as expected reliability / availability requirements for end users, plant owners and / or plant operators.

[0031] A fundamental aspect of the new and useful computer executable program is the processing of signals through function / physics-based models configured to evaluate historical and current specific asset / plant behavior and merge them with their data-driven analysis configured to estimate deviations from expected, daily, operating behavior of the monitored equipment and link such models (and / or their outputs) to related system and failure / deterioration events. Examples of such system and failure / deterioration events include: asset emergency shutdowns and / or start-up failures, environmental and safety impacts, asset performance degradation, plant production loss, asset operability cost increase, material deterioration and component failures.

[0032] The disclosure herein also relates to optimizing maintenance scenarios through modeling of the time to event analysis and risk assessment.

[0033] In one aspect, the present subject-matter is directed to the fact that the model on which the maintenance optimization method is based allows to detect anomalies with respect to the expected behavior, to assign a class and a severity to the anomaly, in order to predict future anomaly trends, and then to calculate the time to schedule and / or perform the related maintenance.

[0034] Anomaly detection can be achieved through a machine learning technique configured to define whether a signal pattern is healthy or unhealthy with respect to a reference pattern. The reference pattern is adjusted thanks to a cohort similarity analysis and can be estimated through physics-based models (e.g. thermodynamic performance or material design properties modeling) and / or data-driven models (clustering of cohort data or historical unit trends). These patterns are computed as numerical features, extracted from a batch of data (size can be adjusted). Anomaly classification is then performed by applying a multinomial regression classifier. The identified anomaly classes are then split into two macro groups: system anomalies and sensor anomalies. These groups are then processed through a risk model that assigns a severity and a likelihood to any event the anomaly should be related to.

[0035] Reference will now be made to the drawings, Figure 1 A maintenance system, globally indicated by the reference 1, is shown, comprising a cohort or group 2 of turbomachinery assets to be maintained, and a control logic unit 3, in which a maintenance optimization method is executed to maintain the turbomachinery assets of the cohort 2. The control logic unit 3 is operatively connected to the cohort 2 (or “complex group”) of turbomachinery assets to be maintained by an internet network 4. It is intended that the system (i.e. the control logic unit 3) can also be implemented in a cloud computing system.

[0036] By way of example, the cohort 2 of turbomachinery assets comprises a plurality of gas turbines. In particular, three gas turbines 21, 22 and 23 are shown. It is clear that the cohort 2 to be maintained can comprise a different number of gas turbines 21, 22 and 23.

[0037] Each gas turbine 21, 22 or 23 comprises the turbomachine itself for the production of energy, and the related auxiliary systems for its operation, such as pumps, actuators, pipes, etc. generally required for the operation of each gas turbine 21, 22 and 23 or turbomachinery asset.

[0038] In some embodiments, each gas turbine 21, 22 and 23 can be equipped with a signal acquisition module, indicated with the references 211, 221 and 231, respectively, each configured to receive detection signals (typically electrical signals) from sensors installed on the gas turbines 21, 22 and 23, and to process said signals, for example filtering and amplifying them, before any further processing of the signals.

[0039] Moreover, each gas turbine 21, 22 and 23 to be maintained also comprises a receiving- sending module 212, 222 and 232 for connecting and sending signals through the Internet network 4, so as to send the signals collected by each gas turbine 21, 22 and 23 of the fleet 2 through it. The signals can also be transmitted from each gas turbine 21, 22 and 23 of the fleet 2 of turbomachinery assets to be maintained to the control logic unit 3 through different channels, for example through radio transmission, optical fiber, etc.

[0040] In some embodiments, and with particular reference to Figure 2 , the central control unit 3 can comprise a processor 31, a bus 32 to which the processor 31 is connected, a database 33 connected to the bus 32 so as to be accessed and controlled by the processor 31, a computer readable memory 34 also connected to the bus 32 so as to be accessed and controlled by the processor 31, a receiving-sending module 35 connected to the bus 32, configured to receive data from and send data to the corresponding receiving-sending modules 212, 222 and 232 of the gas turbines 21, 22 and 23 to be maintained through the Internet network 4.

[0041] In some embodiments, the central control unit 3 can be implemented or embodied as a cloud computing system, a computer network or other device capable of processing data by running appropriate computer programs based on the maintenance optimization method or algorithm.

[0042] The control logic unit 3 is configured to execute one or more computer programs for performing the optimization method for maintaining the fleet 2, which will be better disclosed below.

[0043] In Figure 3 the maintenance optimization method / algorithm is indicated as a whole by the reference 5 and, as mentioned above, is embodied as a computer program, as already mentioned, executed by a computer or any processing device that can be incorporated in the central control unit 3.

[0044] The maintenance optimization method 5 comprises two main processing steps or branches, namely a model setting step 6 (which can also be performed offline) and an online management step 7.

[0045] As mentioned above, the model setting step 6 is usually performed offline, i.e. not in real time, neither during the normal operation of the gas turbines 21, 22 and 23, nor during the set-up phase of the method or at the time of asset installation and / or configuration update, such as major maintenance activities or asset replacement / upgrade. The re-training of the model can be performed manually by an expert or can be triggered by an automatic logic check, for example when the newly computed health feature is above a certain threshold with respect to the initial feature. The offline training step 6 can also be performed on historical fleet data and it saves the configuration parameters CP to be used by the online management step 7, as better disclosed hereinafter. In this respect, the result of the model setting step 6 is a set of configuration parameters CP for configuring the online management step 7 in order to optimize the latter to detect anomalies of the fleet 2 of turbomachinery assets and determine maintenance actions.

[0046] The offline model setting step 6 can be performed at any time when the model is not working properly, so that for example when the model (better described below) based on this method provides no longer acceptable results, it is necessary to re-parameterize the model.

[0047] The online management step 7 is mainly concerned with processing online data, i.e. data collected while the gas turbines 21, 22 and 23 to be maintained are in operation, so it operates when the control logic unit 3 is directly connected to the asset monitoring and communication network. The software processing capacity must be in line with the data throughput to avoid introducing delays in the infrastructure and to ensure that the model results are available in time (even in real time) so as to guarantee the execution of the anomaly early detection and provide maintenance actions sufficiently in advance to plan the required on-site activities.

[0048] Both the above offline model setting step 6 and the online management step 7, implemented in a suitable software language, can be performed in an edge or remote / cloud infrastructure, depending on the implementation of the maintenance system 1. The edge infrastructure is considered suitable only when the network security and encryption requirements are adequately considered by the software OEM.

[0049] In one embodiment, the offline model setting step 6 comprises an input signals sub-step 61 in which a set of input signals from the gas turbines 21, 22 and 23 of the fleet 2 to be monitored and maintained is received. More specifically, the signals are acquired from sensors installed on the assets for checking the operation of the assets and are eventually processed by the control system or other acquisition boards, such as the signal acquisition modules indicated with reference numerals 211, 221 and 231 for each gas turbine 21, 22 and 23, respectively, as mentioned above. Moreover, the input data set comprises all the parameters that can be estimated from the acquired signals by data-driven analysis and / or physics-based models, such as performance, consumption, emissions, material degradation, component failure rate, etc. As mentioned, the processing step can be performed on an edge or cloud infrastructure. The sampling rate of each signal can be different.

[0050] Further, the off-line model setting step 6 comprises a pre-processing sub-step 62 of the signals, in which filtering and / or signal de-correlation processing is performed to define the window of asset operating conditions in which each signal must be processed and / or monitored. For example, signals observed in a certain speed range or power range, or in the case of the engine not running.

[0051] Further, the pre-processing sub-step 62 removes the correlation between the signals of different systems in order to define the expected behavior of the signals independently of the asset specific operating conditions or environmental conditions. These signals will be processed by evaluating the residual between their measured values and their expected values de-correlated. Examples of these residuals are the distance between the estimated material degradation and the asset component aging with respect to the calculated baselines obtained through OEM models and analysis.

[0052] After the pre-processing sub-step 62, the off-line model setting step 6 comprises a similarity analysis or identification sub-step 63 for identifying similarity groups within the asset fleet or group 2. For example, the gas turbines 21, 22 and 23 in the same production line are usually very similar in terms of asset configuration and operating parameters. The similarity analysis is done by using distance or similarity metric clustering techniques.

[0053] After the previous step, a calculation step of health features or fleet or group feature extraction step 64 is performed with respect to the previously identified similarity fleet 2. The extracted features are statistical / mathematical parameters. The list of features to be used depends on the type of signal. Examples of extracted features are median, mean, standard deviation, percentiles, derivative, kurtosis, skewness, projection of the signal onto the principal components and wavelet decomposition components obtained by spectral analysis. The training of the anomaly detection and classification algorithm can be performed by means of supervised and / or unsupervised methods. In this embodiment, an example of the use of a hybrid supervised-unsupervised method is described. In this step, the extracted features are processed using clustering algorithms (for example, k-means or expectation-maximization algorithms, as better explained below) and the obtained clusters are classified into different categories (for example, healthy, step change, noise, asymmetric noise, spike, drift and out-of-range / values) by the OEM experts and / or automatic algorithms based on historical experience and fleet data. If the anomalies identified by the experts are not sufficient to perform the setting of the model, the anomalies are simulated and injected into the dataset and the feature extraction is performed again.

[0054] In some embodiments, the classification algorithm can be based on supervised or unsupervised methods or hybrid methods.

[0055] More specifically, the group feature extraction step 64 comprises a sub-step of identifying time frames in which the assets are operating as expected and computing health features for each identified similarity group of gas turbines 21, 22 and 23. Then the health statistical / mathematical features or parameters within at least one time frame are extracted, time frames in which the assets are operating abnormally are identified, and abnormal features for each identified similarity group of gas turbines 21, 22 and 23 are computed, thereby extracting abnormal statistical / mathematical features or parameters within at least one time frame. If a sufficient number and type of abnormalities are not identified on the signal history trends, then an abnormality is simulated and injected into the signal trends, and the abnormal feature extraction step 64 is performed in the time frame in which the abnormality has been injected.

[0056] The feature reconstruction algorithm (such as Auto-Associative Kernel Regression (AAKR)) is then trained (see step 65) using the clusters of features representative of the health condition, which is able to estimate the expected features of an observation signal by a weighted average of the historical features extracted with respect to the observation of the health signal. Then the anomaly detection is performed by identifying the signal features with residuals above a distance threshold as abnormal. Then a classifier is trained to classify the identified abnormalities with respect to the classes previously selected and identified by the expert. Then a statistical sampling is performed on the health and abnormal features computed within the time frames contained during the entire training and validation to reduce the size of the feature dataset, thereby preserving the same statistical distribution of the health and abnormal features in all signals.

[0057] It can be seen that the anomaly detection is performed dynamically. The threshold of the detection depends on the specific engine condition as it uses a signal reconstruction technique.

[0058] The configuration parameters CP are the output of the offline model setting step 6, which aims at configuring the online anomaly detection step of the online computation step 7, as better disclosed below. These parameters are for example the health and abnormal features, the composition of the similarity asset groups, the signal decorrelation curves, the risk analysis threshold, the anomaly detection and classification model parameters, the fleet statistics parameters. More generally, the configuration parameters CP are constituted by the health and abnormal features extracted within the time frames of the execution of the model setting step 6 and the setting parameters of the identification and classification models trained in the model setting step 6.

[0059] The online computation step 7 comprises an input signals sub-step 71 to define a set of input signals to be considered, in particular to compute one or more additional parameters to assess the health status of the turbomachinery assets 21, 22 and 23, the one or more additional parameters being selected from the group consisting of: performance, emissions, component degradation. This method step is similar to the input signals sub-step 61 of the offline training step 6, or can be identical. The signals in both steps are acquired from sensors installed on the assets of the fleet 2, i.e. in the gas turbines 21, 22 and 23, and possibly processed by a control system or other acquisition board, such as the signal acquisition modules indicated with the above mentioned reference numerals 211, 221 and 231, respectively. Moreover, the input data set comprises all the parameters that can be estimated from the acquired signals, like performance, consumption, emissions, material degradation, component failure rate, etc. The processing step can be performed on the edge or on a cloud infrastructure. Also in this case, the sampling rate of each signal can be different.

[0060] Similarly to the offline training step 6, the online management step 7 comprises a pre-processing sub-step 72 of the signals, to define the window of asset operating conditions that must be processed and / or monitored for each signal. For example, there are signals that need to be observed within a certain speed range or power range, or in the case of engine not running. Moreover, this sub-step removes the correlation between the signals of different systems, in order to define the expected behaviour of the signals independently from the asset specific operating conditions or environmental conditions. These signals will be processed by evaluating the residual between their measured values and their de-correlated expected values.

[0061] In some embodiments, the sub-steps 61 and 62 of the offline model setting step 6 and the sub-steps 71 and 72 of the online management step 7 can be identical to each other and not distinguished respectively.

[0062] The configuration parameters (anomaly detection validation) are defined by comparison between healthy and unhealthy conditions identified on the historical fleet data. The configuration parameters CP are extracted on time series, so the validation is performed by comparing the numerical features extracted on time series.

[0063] By way of example, Figure 4 A set of curves 80 is shown, each curve representing an operating signal acquired from one of the sensors of one of the gas turbines 21, 22 and 23, suitably de-correlated and thus output from the pre-processing sub-step 72, as described above. In particular, Figure 4 Six plotted curves are shown, indicated with the reference numerals 81, 82, 83, 84, 85 and 86, respectively.

[0064] By visual inspection only Figure 4It is evident that curves 84, 85 and 86 all oscillate around the horizontal axis X (i.e. the abscissa) without significant deviation. Curve 83 shows that, in a first portion, it deviates from the X axis and therefore, if compared with the other curves 84, 85 and 86, shows anomalous behaviour; then, as time passes, the signal according to this curve 83 returns to oscillate around the X axis. This can mean that the monitored sensor has a temporary phase at the beginning of the operation, after which it remains correctly operating. Finally, curves 81 and 82 significantly deviate from the X axis along time, indicating that they represent individual sensors whose behaviour has anomalous operation with respect to those represented by curves 84, 85 and 86.

[0065] Downstream of the pre-processing sub-step 72 of the signals received from the fleet 2, a statistical feature extraction 73 is provided, which extracts the features required for the operation of the anomaly detection and classification step, which will be better described hereinafter.

[0066] The extracted features are statistical / mathematical parameters. The list of features to be used depends on the type of signal analysed. Examples of features extracted by the statistical / mathematical feature extraction step 73 are median, mean, standard deviation, percentiles, derivative, kurtosis, skewness, projection on principal components and wavelet decomposition components and combinations thereof. The statistical data are extracted in one or more defined time frames, e.g. a week, a month, etc., the sampling rate being available from the acquisition system / sensor or defined by the OEM expert.

[0067] Then, an anomaly detection sub-step 74 is performed, which identifies whether the signals or set of signals received from the gas turbines 21, 22 and 23 of the fleet 2 are anomalous or not. In some embodiments, the method can be performed by machine learning algorithms, implementing supervised and / or unsupervised methods. The anomaly detection aims at identifying which signals have an anomalous pattern of features with respect to the pattern of healthy features. This detection can be done by using signal reconstruction techniques, such as Auto-Associative Kernel Regression (AAKR), in which case the expected signal features are reconstructed with respect to the pattern of healthy features available as configuration parameters. The comparison between the reconstructed features and the measured features is performed by using distance metrics or similarity metrics (likelihoods) and comparing them with a threshold, as better explained hereinafter.

[0068] It can be seen that the online computation step 7 performs anomaly detection and classification. Moreover, it is based on a multivariate analysis of the different signal anomalies, increasing the assessment of the severity of the anomaly by discriminating between system and sensor faults. System anomalies usually have different features from sensor anomalies and can involve more than one signal. The algorithm is able to identify several classes of anomalies of system and sensor behaviour.

[0069] In order to calculate its output by means of the above algorithm, the anomaly detection sub-step 74 uses the model configuration parameters CP generated and received by the anomaly detection and classification training sub-step 65 of the offline training step 6 and the team feature extraction sub-step 64.

[0070] After detecting the anomaly between the signals as disclosed above, there is an anomaly classification sub-step 75. This sub-step 75 identifies whether the signal or set of signals from the sensors has an anomaly. This method can be performed by a supervised method, adjusted by an unsupervised method.

[0071] The detection and classification algorithm is applied to all the signals, not to individual signals. The signals acquired online from the dynamic system are defined time series. This algorithm processes all the time series acquired from the asset in all operating conditions, such as steady state, transients and engine not running conditions.

[0072] Moreover, all the subsystems of the asset are monitored by means of anomaly detection and classification. Furthermore, according to the prediction of the future signal behaviour, a risk assessment is performed to understand which is the right timeline for the maintenance execution.

[0073] The supervised method can be performed by determining the characteristic reference pattern of the anomaly class that must be detected. The anomaly class can be selected from the historical data according to the expert experience or simulated. Then the anomaly identification is performed by means of techniques such as logistic regression, K-NN or Bayesian networks, respectively using distance metrics or similarity metrics (likelihoods) and, in this case, also comparing them with a threshold. In the distance metrics, the Euclidean and Wesserstein metrics can be mentioned. Moreover, the implementation of the techniques is made based on the signal being analysed and processed.

[0074] Generally, the Wesserstein metric distance is the most common because it is mainly used for time series; while for the performance calculation parameters, the Euclidean metric distance is usually preferred. As for the similarity metrics, they are usually applied to identify the similarity groups between the turbomachinery assets 21, 22 and 23 for each signal.

[0075] The logistic regression technique is the most used in the current model because it matches the current performance and accuracy requirements better than the other available techniques.

[0076] The identified classes are different anomaly types. Here, by way of example, a set of classes that can be identified is listed: signal frozen, signal drift, step change, symmetric noise, asymmetric noise, spike and abnormal range.

[0077] An unsupervised method can be executed to check the accuracy of the classifier periodically in order to determine if the abnormal classes are stable and / or if new abnormal types must be added to the list of classes. Clustering will be performed on the features extracted within the last time frame (the number of time frames can be arbitrary). If the clusters are centered with respect to the abnormal / healthy clusters assigned during the model setup, the model is stable and no update is needed; otherwise a new model setup will be performed. There are several clustering techniques that can be used. In particular, in some embodiments, a maximum likelihood clustering technique is used, which is often applied because it better adapts to the statistical distribution of the signals received and to be processed, thus optimizing the model accuracy, thereby reducing the overall error. In other embodiments, a k-means clustering technique can also be applied to signals with similar and symmetric statistical distribution.

[0078] The identified abnormal classes are then associated with the assets they refer to; in particular, the classification will lead to the assignment of the abnormality to a sensor fault or a system / asset fault. The assignment will be performed by a logic flowchart defined by the OEM experts. For example, for a certain signal, the abnormality classified as signal frozen or noise is assigned to a sensor fault; while the abnormality classified as drift is assigned to a system fault. Then the maintenance optimization method 5 will be executed in order to determine the maintenance actions to be performed. Figure 3 The two macro abnormal classes are treated differently in the flowchart shown.

[0079] In other words, in this embodiment, the abnormal classification sub-step 75 can have two different outcomes in case of a system abnormality (drift or abnormal range / values) or a sensor abnormality (such as frozen, noise, etc.). In the described embodiment, for each of the two possible outcomes, a different handling procedure is executed. In this way, the maintenance optimization method 5 allows to discriminate between faults or abnormalities for a corresponding discrimination of the maintenance treatment and the maintenance operations to be performed.

[0080] In particular, in case the abnormal classification 75 detects a system abnormality, a risk assessment step 761 is executed. This risk assessment step 761 estimates the risk of any event that requires a maintenance task to be performed at unit (i.e. turbomachinery 21 or 22 or 23 of the illustrated embodiment) shutdown. There are several categories of events related to the gas turbines 21, 22 and 23: trip, engine shutdown, performance degradation, unscheduled maintenance and unscheduled engine / component replacement. The risk of each event is assessed in terms of probability and impact.

[0081] The risk assessment step 761 combines the risks of system failure, asset deterioration and component / material aging. The risk assessment is done by evaluating these phenomena according to their impact on the previously listed asset unavailability events. The weighting and combination criteria used to perform the risk assessment are based on OEM experience and design models. The risk model can be physics-based, data-driven or hybrid type and its development, validation and updating are based on test data, online monitoring data and inspection results / measures.

[0082] The inputs of the risk assessment step 761 software block or module are the signals identified as affected by the system anomaly, all other signals needed to run the risk assessment and the signals affected by the sensor anomaly in the range / dates of influence of the maintenance.

[0083] Reference is now made to Figure 5 and also to Figure 4 It can be seen that, after processing the signals received from the gas turbines 21, 22 and 23, an example of the results of the risk assessment sub-step 761. More specifically, it can be seen from the signals 84, 85 and 86 that no anomaly has been detected, so the sensors that generated this signal have not presented a maintenance problem. On the contrary, the risk assessment sub-step 761 generates the curves 81', 82' and 83' corresponding to the curves 81, 82 and 83, which, as mentioned above, deviate from the behavior of the other curves, in particular, for the present embodiment, from the X axis. If the curves 81', 82' and 83' reinforce the different behavior of the other curves compared to the other curves.

[0084] In this way, the anomalous behavior of the curve 83' is highlighted in the first part of the detection time frame. In the same way, the anomalous behavior of the curve 84' and the curve 85' can be embodied in the last part of the detection time frame.

[0085] Then, after the risk assessment step 761, the model prediction and maintenance date sub-step 762 is executed, which estimates the maintenance time based on the predictive analysis of each model and its combination. The prediction can be done by time series prediction techniques or probabilistic prediction based on aging parameters. The maintenance date is estimated by calculating the total predicted risk expected to reach a certain limit value.

[0086] Finally, the maintenance action sub-step 78 is executed, whereby the actions to be performed, both at unit shutdown and / or while the unit is running, are listed and displayed on a monitor (not shown in the figure) or are generally provided to the user by any other suitable device for listing the appropriate maintenance operations.

[0087] More specifically, the maintenance action step 78 defines the maintenance and / or risk mitigation actions to be performed, distinguishing between those that can be performed while the unit is running and those to be performed at the next shutdown.

[0088] The list of maintenance actions to be performed is based on the identified anomalies and risk assessment, and the maintenance treatment is prioritized with respect to the ranking / contribution in the risk assessment of its related system / asset anomaly.

[0089] In some embodiments, the mapping between anomalies / aging / deterioration phenomena and appropriate maintenance actions is established based on OEM experience and maintenance policies.

[0090] In case of anomaly classification 75 detecting one (or at least one) sensor failure, a severity assignment 771 step is performed, in which a severity is assigned to the anomaly identified as a sensor failure. The severity assignment 771 assigns a severity score to the anomaly classified as a sensor failure based on the anomaly type and duration and sensor redundancy. Examples of anomalies can be, for example, signal freeze, noise and spikes, which have different priorities: a freeze is considered more severe than noise because it leads to a missing monitoring of the item. Also, the frequency and duration of the anomaly will be considered to assess the severity. The severity can be a percentage value or an integer value range (0-10) or any other ranking system capable of prioritizing the event severity.

[0091] Finally, it is checked whether the result of the severity assignment step 771 impacts the maintenance scope or date (see sub-step 772). More specifically, in this step, it is mapped if a correction or risk mitigation action can be performed online or if an engine shutdown is required. The mapping function depends on the anomaly type, signal type and severity, from OEM experience and maintenance policies.

[0092] If said result of the severity assignment step 771 impacts the maintenance scope or date, the risk assessment step 761 already described above is performed. Otherwise, if there is no impact, the maintenance treatment step 68 is performed, also already described above.

[0093] The maintenance optimization method / algorithm 5 is able to process all the turbine monitoring signals of the entire fleet. Then the maintenance optimization method / algorithm 5 starts from the processing of the input time series.

[0094] Also, the maintenance optimization method / algorithm 5 is a predictive maintenance solution. The maintenance plan is dynamic as it is based on the evolution of the signals over time.

[0095] The advantage of the maintenance optimization method 5 is that it combines the OEM maintenance service knowledge and the digital service provider capabilities. Technically, this is done by combining the functional / physical knowledge of the failure mode modeling and the data-driven analytics developed based on the fleet monitoring data and inspection experience.

[0096] The commercial advantage is the merging of the OEM physical model with data driven analytics allows to optimize maintenance scenarios versus customer requirements: maximum reliability / availability, minimization of interruption duration, optimization of the interruption scope of the job, maximization of the maintenance cycle time, minimization of the risk of critical failures. This flexibility is the advantage of the OEM asset owner that has the technical design, the online monitoring data, as well as the historical failure statistics and inspection results.

[0097] Figure 3 The various arrangements and steps shown in the figures can be combined in various ways. In particular, some common steps of the offline model setup step 6 and the online computation step 7 can be performed by the same software block or module.

[0098] While aspects of the present application have been described with reference to various specific embodiments, it will be apparent to those of ordinary skill in the art that numerous modifications, variations, and alternatives can be devised in light of the above teachings without departing from the spirit and scope of the present disclosure. Further, unless otherwise indicated herein, the order or sequence of any process or method steps can be varied or re-sequenced in alternative embodiments.

[0099] Reference will now be made in detail to implementations of the present disclosure, one or more examples of which are illustrated in the drawings. Each example is provided by way of explanation of the present disclosure, not limitation of the present disclosure. In fact, it will be apparent to those skilled in the art that various modifications and variations can be made in the implementations disclosed without departing from the scope or spirit of the present disclosure. References in this specification to “one implementation” or “an implementation” or “some implementations” mean that a particular feature, structure, or characteristic described in connection with the implementation is included in at least one implementation of the disclosed subject matter. Thus, the appearances of the phrase “in one implementation” or “in an implementation” or “in some implementations” in various places in the specification are not necessarily referring to the same implementation. Further, the terms “comprise,” “comprising,” “include,” “including,” and “has,” “having,” their variants, and / or the like are intended to be open-ended terms i.e., a process, method, object, or the like that comprises or includes an item does not exclude additional items. In some embodiments, the terms “comprise,” “comprising,” “include,” “including,” and “has,” “having,” their variants, and / or the like can be used interchangeably with “consist of,” “consisting of,” “consists in,” and / or “consists of.”

[0100] When introducing elements of various embodiments, the articles “a,” “an,” “the,” and “said” are intended to mean that there are one or more of the elements. The terms “comprise,” “include,” and “have” and variations thereof are intended to be inclusive and are meant to be open-ended and allow for additions.

Claims

1. A computer-implemented method for maintenance optimization of a fleet or a group of plants, wherein the plants comprise one or more turbomachinery assets, wherein each turbomachinery asset comprises a turbomachine and related auxiliary systems, and wherein each of the turbomachinery assets comprises one or more installed sensors, each capable of generating an operating signal of the turbomachinery asset, the method comprising the steps of: - computing model configuration parameters by comparing healthy and unhealthy conditions of turbomachinery assets found in historical fleet data; - acquiring, within a defined time frame, data consisting of operating signals from the sensors of the turbomachinery assets; - filtering and decorrelating the operating signals to remove correlations between different system operating signals, resulting in decorrelated signals; - extracting features from the decorrelated signals; - comparing the features to the model configuration parameters using a multivariate problem approach to identify one or more anomalies in the features; - classifying the anomalies as system anomalies related to the turbomachinery assets of the fleet or as sensor anomalies, wherein, if the classifying step identifies one or more system anomalies, then performing a risk assessment step for estimating the risk of any event requiring a maintenance task and the appropriate time for performing the maintenance, otherwise, if the classifying step detects one or more sensor anomalies, then performing a severity assessment step for assigning a severity to the anomalies identified as sensor faults; and allocating and sending a treatment to an end user.

2. The method of claim 1, wherein the anomaly detection step is performed by machine learning techniques in a supervised and / or unsupervised approach, wherein in the supervised approach the anomaly detection is performed by comparing the features extracted on the acquired / calculated parameters to healthy reference signal features.

3. The method of claim 1, wherein the features extracted on the acquired / calculated parameters are compared to healthy reference signal features by using signal reconstruction techniques.

4. The method of claim 3, wherein the signal reconstruction techniques are performed by Auto-Associative Kernel Regression (AAKR) techniques, wherein the anomaly is identified when the distance between the reconstructed features and the healthy features is higher than a correlation threshold, the correlation threshold being one of the model configuration parameters, wherein the metric used for feature comparison is the Euclidean distance or the Wesserstein distance or a likelihood metric.

5. The method of claim 1, wherein the anomaly classification step is performed by a classifier determining the class of the anomaly detected at the detection step, wherein the anomaly classes are selected from historical data or by expert knowledge.

6. The method of claim 5, wherein the anomalies are classified in the anomaly classification step into different classes selected from the following: signal freeze, step change, symmetric noise, asymmetric noise, spike, abnormal or anomalous range, drift; and wherein the identified abnormality classes are then associated with the turbomachinery assets of the fleet they refer to, so that the classification allows to assign the abnormality to a sensor fault or system fault / degradation related to the turbomachinery assets of the fleet.

7. The method of claim 1, wherein the acquisition step comprises the steps of: receiving input signals from the turbomachinery assets of the fleet to be monitored and maintained, computing one or more additional parameters to assess the health condition of the turbomachinery assets, the one or more additional parameters being selected from: performance, emissions, component degradation; and pre-processing the signals received from the turbomachinery assets and the other additional parameters computed to assess the health condition of the turbomachinery assets, wherein filtering and / or signal decorrelation processing is performed to obtain normalized signals with respect to specific asset operating conditions.

8. The method of claim 1, wherein the risk assessment step estimates the risk of any event requiring the execution of a maintenance task at one of the turbomachinery assets, and wherein the method comprises a prediction step to estimate the maintenance time by means of a prediction analysis based on time series prediction techniques or probabilistic prediction based on aging parameter extrapolation.

9. The method of claim 1, wherein a severity assignment step is performed to assign a severity score to the abnormality classified as a sensor fault based on abnormality type and duration and sensor redundancy.

10. The method of claim 1, comprising a step of determining whether a sensor fault affects the maintenance scope or date, wherein if the result of the severity assignment step performed with respect to the sensor abnormality affects the maintenance scope or date, then the risk assessment step is performed.

11. The method of claim 10, wherein the event requiring the execution of a maintenance task at one of the turbomachinery assets comprises an event reducing asset availability, including at least one of: trip, unscheduled downtime, system fault, asset degradation and component / material aging, performance drop, engine / component failure.

12. The method of claim 1, wherein the model configuration step for computing model configuration parameters is performed offline.

13. The method of claim 1, wherein the model configuration step comprises the steps of: receiving input signals from the turbomachinery assets of the fleet to be monitored and maintained; computing any parameters needed to describe the health condition of the assets, including performance, emissions, component degradation; and pre-processing the signals received from the turbomachinery assets and the computed parameters, wherein filtering and / or signal decorrelation processing is performed to obtain normalized signals with respect to specific asset operating conditions.

14. The method of claim 13, wherein the model configuration step comprises the steps of: identifying similarity groups among the turbomachinery assets of the fleet, performed by using clustering techniques of distance or similarity measures on the operating signals and parameters of all the assets of the fleet; and performing a group feature extraction step, comprising the sub-steps of: identifying similarity groups among the turbomachinery assets of the fleet, performed by using clustering techniques of distance or similarity measures on the operating signals and parameters of all the assets of the fleet; and performing a group feature extraction step, comprising the sub-steps of: identifying time frames where the asset is operating as expected and computing health features on each of the identified similarity groups of the turbomachinery asset, extracting health statistical / mathematical features or parameters over at least one time frame, identifying time frames where the asset is operating abnormally and computing abnormal features on each of the identified similarity groups of the turbomachinery asset, extracting abnormal statistical / mathematical features or parameters over at least one time frame, wherein in case a sufficient number and type of abnormalities are not identified on the signal historical trends, simulating the abnormality and injecting the abnormality into the signal trends, and performing the abnormal feature extraction step between the time frames where the abnormality has been injected; and training the abnormality identification and classification steps, the identification and the classification steps being respectively a signal reconstruction AAKR and a multinomial classification model; wherein the configuration parameters consist of the health and abnormal features extracted over the time frames where the model configuration step is performed and the setting parameters of the identification and classification models trained within the model configuration step.

15. The method of claim 14, wherein the statistical / mathematical features or parameters are median, mean, standard deviation, percentile, derivative, kurtosis, skewness, projection on principal components and wavelet decomposition components obtained by spectral analysis, and combinations thereof.

Images